fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "free-space-tree.h"
  37 #include "math.h"
  38 #include "sysfs.h"
  39 #include "qgroup.h"
  40
  41 #undef SCRAMBLE_DELAYED_REFS
  42
  43 /*
  44  * control flags for do_chunk_alloc's force field
  45  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  46  * if we really need one.
  47  *
  48  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  49  * if we have very few chunks already allocated.  This is
  50  * used as part of the clustering code to help make sure
  51  * we have a good pool of storage to cluster in, without
  52  * filling the FS with empty chunks
  53  *
  54  * CHUNK_ALLOC_FORCE means it must try to allocate one
  55  *
  56  */
  57 enum {
  58         CHUNK_ALLOC_NO_FORCE = 0,
  59         CHUNK_ALLOC_LIMITED = 1,
  60         CHUNK_ALLOC_FORCE = 2,
  61 };
  62
  63 static int update_block_group(struct btrfs_trans_handle *trans,
  64                               struct btrfs_root *root, u64 bytenr,
  65                               u64 num_bytes, int alloc);
  66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  67                                 struct btrfs_root *root,
  68                                 struct btrfs_delayed_ref_node *node, u64 parent,
  69                                 u64 root_objectid, u64 owner_objectid,
  70                                 u64 owner_offset, int refs_to_drop,
  71                                 struct btrfs_delayed_extent_op *extra_op);
  72 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  73                                     struct extent_buffer *leaf,
  74                                     struct btrfs_extent_item *ei);
  75 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  76                                       struct btrfs_root *root,
  77                                       u64 parent, u64 root_objectid,
  78                                       u64 flags, u64 owner, u64 offset,
  79                                       struct btrfs_key *ins, int ref_mod);
  80 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  81                                      struct btrfs_root *root,
  82                                      u64 parent, u64 root_objectid,
  83                                      u64 flags, struct btrfs_disk_key *key,
  84                                      int level, struct btrfs_key *ins);
  85 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  86                           struct btrfs_root *extent_root, u64 flags,
  87                           int force);
  88 static int find_next_key(struct btrfs_path *path, int level,
  89                          struct btrfs_key *key);
  90 static void dump_space_info(struct btrfs_fs_info *fs_info,
  91                             struct btrfs_space_info *info, u64 bytes,
  92                             int dump_block_groups);
  93 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
  94                                     u64 ram_bytes, u64 num_bytes, int delalloc);
  95 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
  96                                      u64 num_bytes, int delalloc);
  97 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  98                                u64 num_bytes);
  99 int btrfs_pin_extent(struct btrfs_root *root,
 100                      u64 bytenr, u64 num_bytes, int reserved);
 101 static int __reserve_metadata_bytes(struct btrfs_root *root,
 102                                     struct btrfs_space_info *space_info,
 103                                     u64 orig_bytes,
 104                                     enum btrfs_reserve_flush_enum flush);
 105 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 106                                      struct btrfs_space_info *space_info,
 107                                      u64 num_bytes);
 108 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 109                                      struct btrfs_space_info *space_info,
 110                                      u64 num_bytes);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED ||
 117                 cache->cached == BTRFS_CACHE_ERROR;
 118 }
 119
 120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 121 {
 122         return (cache->flags & bits) == bits;
 123 }
 124
 125 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 126 {
 127         atomic_inc(&cache->count);
 128 }
 129
 130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 131 {
 132         if (atomic_dec_and_test(&cache->count)) {
 133                 WARN_ON(cache->pinned > 0);
 134                 WARN_ON(cache->reserved > 0);
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_root *root,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&root->fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE);
 232         set_extent_bits(&root->fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_root *root,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&root->fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE);
 247         clear_extent_bits(&root->fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_root *root,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(root, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(root->fs_info, cache->key.objectid,
 271                                        bytenr, 0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(root, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (!cache->caching_ctl) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         ctl = cache->caching_ctl;
 320         atomic_inc(&ctl->count);
 321         spin_unlock(&cache->lock);
 322         return ctl;
 323 }
 324
 325 static void put_caching_control(struct btrfs_caching_control *ctl)
 326 {
 327         if (atomic_dec_and_test(&ctl->count))
 328                 kfree(ctl);
 329 }
 330
 331 #ifdef CONFIG_BTRFS_DEBUG
 332 static void fragment_free_space(struct btrfs_root *root,
 333                                 struct btrfs_block_group_cache *block_group)
 334 {
 335         u64 start = block_group->key.objectid;
 336         u64 len = block_group->key.offset;
 337         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 338                 root->nodesize : root->sectorsize;
 339         u64 step = chunk << 1;
 340
 341         while (len > chunk) {
 342                 btrfs_remove_free_space(block_group, start, chunk);
 343                 start += step;
 344                 if (len < step)
 345                         len = 0;
 346                 else
 347                         len -= step;
 348         }
 349 }
 350 #endif
 351
 352 /*
 353  * this is only called by cache_block_group, since we could have freed extents
 354  * we need to check the pinned_extents for any extents that can't be used yet
 355  * since their free space will be released as soon as the transaction commits.
 356  */
 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 358                        struct btrfs_fs_info *info, u64 start, u64 end)
 359 {
 360         u64 extent_start, extent_end, size, total_added = 0;
 361         int ret;
 362
 363         while (start < end) {
 364                 ret = find_first_extent_bit(info->pinned_extents, start,
 365                                             &extent_start, &extent_end,
 366                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 367                                             NULL);
 368                 if (ret)
 369                         break;
 370
 371                 if (extent_start <= start) {
 372                         start = extent_end + 1;
 373                 } else if (extent_start > start && extent_start < end) {
 374                         size = extent_start - start;
 375                         total_added += size;
 376                         ret = btrfs_add_free_space(block_group, start,
 377                                                    size);
 378                         BUG_ON(ret); /* -ENOMEM or logic error */
 379                         start = extent_end + 1;
 380                 } else {
 381                         break;
 382                 }
 383         }
 384
 385         if (start < end) {
 386                 size = end - start;
 387                 total_added += size;
 388                 ret = btrfs_add_free_space(block_group, start, size);
 389                 BUG_ON(ret); /* -ENOMEM or logic error */
 390         }
 391
 392         return total_added;
 393 }
 394
 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 396 {
 397         struct btrfs_block_group_cache *block_group;
 398         struct btrfs_fs_info *fs_info;
 399         struct btrfs_root *extent_root;
 400         struct btrfs_path *path;
 401         struct extent_buffer *leaf;
 402         struct btrfs_key key;
 403         u64 total_found = 0;
 404         u64 last = 0;
 405         u32 nritems;
 406         int ret;
 407         bool wakeup = true;
 408
 409         block_group = caching_ctl->block_group;
 410         fs_info = block_group->fs_info;
 411         extent_root = fs_info->extent_root;
 412
 413         path = btrfs_alloc_path();
 414         if (!path)
 415                 return -ENOMEM;
 416
 417         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 418
 419 #ifdef CONFIG_BTRFS_DEBUG
 420         /*
 421          * If we're fragmenting we don't want to make anybody think we can
 422          * allocate from this block group until we've had a chance to fragment
 423          * the free space.
 424          */
 425         if (btrfs_should_fragment_free_space(extent_root, block_group))
 426                 wakeup = false;
 427 #endif
 428         /*
 429          * We don't want to deadlock with somebody trying to allocate a new
 430          * extent for the extent root while also trying to search the extent
 431          * root to add free space.  So we skip locking and search the commit
 432          * root, since its read-only
 433          */
 434         path->skip_locking = 1;
 435         path->search_commit_root = 1;
 436         path->reada = READA_FORWARD;
 437
 438         key.objectid = last;
 439         key.offset = 0;
 440         key.type = BTRFS_EXTENT_ITEM_KEY;
 441
 442 next:
 443         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 444         if (ret < 0)
 445                 goto out;
 446
 447         leaf = path->nodes[0];
 448         nritems = btrfs_header_nritems(leaf);
 449
 450         while (1) {
 451                 if (btrfs_fs_closing(fs_info) > 1) {
 452                         last = (u64)-1;
 453                         break;
 454                 }
 455
 456                 if (path->slots[0] < nritems) {
 457                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 458                 } else {
 459                         ret = find_next_key(path, 0, &key);
 460                         if (ret)
 461                                 break;
 462
 463                         if (need_resched() ||
 464                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 465                                 if (wakeup)
 466                                         caching_ctl->progress = last;
 467                                 btrfs_release_path(path);
 468                                 up_read(&fs_info->commit_root_sem);
 469                                 mutex_unlock(&caching_ctl->mutex);
 470                                 cond_resched();
 471                                 mutex_lock(&caching_ctl->mutex);
 472                                 down_read(&fs_info->commit_root_sem);
 473                                 goto next;
 474                         }
 475
 476                         ret = btrfs_next_leaf(extent_root, path);
 477                         if (ret < 0)
 478                                 goto out;
 479                         if (ret)
 480                                 break;
 481                         leaf = path->nodes[0];
 482                         nritems = btrfs_header_nritems(leaf);
 483                         continue;
 484                 }
 485
 486                 if (key.objectid < last) {
 487                         key.objectid = last;
 488                         key.offset = 0;
 489                         key.type = BTRFS_EXTENT_ITEM_KEY;
 490
 491                         if (wakeup)
 492                                 caching_ctl->progress = last;
 493                         btrfs_release_path(path);
 494                         goto next;
 495                 }
 496
 497                 if (key.objectid < block_group->key.objectid) {
 498                         path->slots[0]++;
 499                         continue;
 500                 }
 501
 502                 if (key.objectid >= block_group->key.objectid +
 503                     block_group->key.offset)
 504                         break;
 505
 506                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 507                     key.type == BTRFS_METADATA_ITEM_KEY) {
 508                         total_found += add_new_free_space(block_group,
 509                                                           fs_info, last,
 510                                                           key.objectid);
 511                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 512                                 last = key.objectid +
 513                                         fs_info->tree_root->nodesize;
 514                         else
 515                                 last = key.objectid + key.offset;
 516
 517                         if (total_found > CACHING_CTL_WAKE_UP) {
 518                                 total_found = 0;
 519                                 if (wakeup)
 520                                         wake_up(&caching_ctl->wait);
 521                         }
 522                 }
 523                 path->slots[0]++;
 524         }
 525         ret = 0;
 526
 527         total_found += add_new_free_space(block_group, fs_info, last,
 528                                           block_group->key.objectid +
 529                                           block_group->key.offset);
 530         caching_ctl->progress = (u64)-1;
 531
 532 out:
 533         btrfs_free_path(path);
 534         return ret;
 535 }
 536
 537 static noinline void caching_thread(struct btrfs_work *work)
 538 {
 539         struct btrfs_block_group_cache *block_group;
 540         struct btrfs_fs_info *fs_info;
 541         struct btrfs_caching_control *caching_ctl;
 542         struct btrfs_root *extent_root;
 543         int ret;
 544
 545         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 546         block_group = caching_ctl->block_group;
 547         fs_info = block_group->fs_info;
 548         extent_root = fs_info->extent_root;
 549
 550         mutex_lock(&caching_ctl->mutex);
 551         down_read(&fs_info->commit_root_sem);
 552
 553         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 554                 ret = load_free_space_tree(caching_ctl);
 555         else
 556                 ret = load_extent_tree_free(caching_ctl);
 557
 558         spin_lock(&block_group->lock);
 559         block_group->caching_ctl = NULL;
 560         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 561         spin_unlock(&block_group->lock);
 562
 563 #ifdef CONFIG_BTRFS_DEBUG
 564         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 565                 u64 bytes_used;
 566
 567                 spin_lock(&block_group->space_info->lock);
 568                 spin_lock(&block_group->lock);
 569                 bytes_used = block_group->key.offset -
 570                         btrfs_block_group_used(&block_group->item);
 571                 block_group->space_info->bytes_used += bytes_used >> 1;
 572                 spin_unlock(&block_group->lock);
 573                 spin_unlock(&block_group->space_info->lock);
 574                 fragment_free_space(extent_root, block_group);
 575         }
 576 #endif
 577
 578         caching_ctl->progress = (u64)-1;
 579
 580         up_read(&fs_info->commit_root_sem);
 581         free_excluded_extents(fs_info->extent_root, block_group);
 582         mutex_unlock(&caching_ctl->mutex);
 583
 584         wake_up(&caching_ctl->wait);
 585
 586         put_caching_control(caching_ctl);
 587         btrfs_put_block_group(block_group);
 588 }
 589
 590 static int cache_block_group(struct btrfs_block_group_cache *cache,
 591                              int load_cache_only)
 592 {
 593         DEFINE_WAIT(wait);
 594         struct btrfs_fs_info *fs_info = cache->fs_info;
 595         struct btrfs_caching_control *caching_ctl;
 596         int ret = 0;
 597
 598         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 599         if (!caching_ctl)
 600                 return -ENOMEM;
 601
 602         INIT_LIST_HEAD(&caching_ctl->list);
 603         mutex_init(&caching_ctl->mutex);
 604         init_waitqueue_head(&caching_ctl->wait);
 605         caching_ctl->block_group = cache;
 606         caching_ctl->progress = cache->key.objectid;
 607         atomic_set(&caching_ctl->count, 1);
 608         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 609                         caching_thread, NULL, NULL);
 610
 611         spin_lock(&cache->lock);
 612         /*
 613          * This should be a rare occasion, but this could happen I think in the
 614          * case where one thread starts to load the space cache info, and then
 615          * some other thread starts a transaction commit which tries to do an
 616          * allocation while the other thread is still loading the space cache
 617          * info.  The previous loop should have kept us from choosing this block
 618          * group, but if we've moved to the state where we will wait on caching
 619          * block groups we need to first check if we're doing a fast load here,
 620          * so we can wait for it to finish, otherwise we could end up allocating
 621          * from a block group who's cache gets evicted for one reason or
 622          * another.
 623          */
 624         while (cache->cached == BTRFS_CACHE_FAST) {
 625                 struct btrfs_caching_control *ctl;
 626
 627                 ctl = cache->caching_ctl;
 628                 atomic_inc(&ctl->count);
 629                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 630                 spin_unlock(&cache->lock);
 631
 632                 schedule();
 633
 634                 finish_wait(&ctl->wait, &wait);
 635                 put_caching_control(ctl);
 636                 spin_lock(&cache->lock);
 637         }
 638
 639         if (cache->cached != BTRFS_CACHE_NO) {
 640                 spin_unlock(&cache->lock);
 641                 kfree(caching_ctl);
 642                 return 0;
 643         }
 644         WARN_ON(cache->caching_ctl);
 645         cache->caching_ctl = caching_ctl;
 646         cache->cached = BTRFS_CACHE_FAST;
 647         spin_unlock(&cache->lock);
 648
 649         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 650                 mutex_lock(&caching_ctl->mutex);
 651                 ret = load_free_space_cache(fs_info, cache);
 652
 653                 spin_lock(&cache->lock);
 654                 if (ret == 1) {
 655                         cache->caching_ctl = NULL;
 656                         cache->cached = BTRFS_CACHE_FINISHED;
 657                         cache->last_byte_to_unpin = (u64)-1;
 658                         caching_ctl->progress = (u64)-1;
 659                 } else {
 660                         if (load_cache_only) {
 661                                 cache->caching_ctl = NULL;
 662                                 cache->cached = BTRFS_CACHE_NO;
 663                         } else {
 664                                 cache->cached = BTRFS_CACHE_STARTED;
 665                                 cache->has_caching_ctl = 1;
 666                         }
 667                 }
 668                 spin_unlock(&cache->lock);
 669 #ifdef CONFIG_BTRFS_DEBUG
 670                 if (ret == 1 &&
 671                     btrfs_should_fragment_free_space(fs_info->extent_root,
 672                                                      cache)) {
 673                         u64 bytes_used;
 674
 675                         spin_lock(&cache->space_info->lock);
 676                         spin_lock(&cache->lock);
 677                         bytes_used = cache->key.offset -
 678                                 btrfs_block_group_used(&cache->item);
 679                         cache->space_info->bytes_used += bytes_used >> 1;
 680                         spin_unlock(&cache->lock);
 681                         spin_unlock(&cache->space_info->lock);
 682                         fragment_free_space(fs_info->extent_root, cache);
 683                 }
 684 #endif
 685                 mutex_unlock(&caching_ctl->mutex);
 686
 687                 wake_up(&caching_ctl->wait);
 688                 if (ret == 1) {
 689                         put_caching_control(caching_ctl);
 690                         free_excluded_extents(fs_info->extent_root, cache);
 691                         return 0;
 692                 }
 693         } else {
 694                 /*
 695                  * We're either using the free space tree or no caching at all.
 696                  * Set cached to the appropriate value and wakeup any waiters.
 697                  */
 698                 spin_lock(&cache->lock);
 699                 if (load_cache_only) {
 700                         cache->caching_ctl = NULL;
 701                         cache->cached = BTRFS_CACHE_NO;
 702                 } else {
 703                         cache->cached = BTRFS_CACHE_STARTED;
 704                         cache->has_caching_ctl = 1;
 705                 }
 706                 spin_unlock(&cache->lock);
 707                 wake_up(&caching_ctl->wait);
 708         }
 709
 710         if (load_cache_only) {
 711                 put_caching_control(caching_ctl);
 712                 return 0;
 713         }
 714
 715         down_write(&fs_info->commit_root_sem);
 716         atomic_inc(&caching_ctl->count);
 717         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 718         up_write(&fs_info->commit_root_sem);
 719
 720         btrfs_get_block_group(cache);
 721
 722         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 723
 724         return ret;
 725 }
 726
 727 /*
 728  * return the block group that starts at or after bytenr
 729  */
 730 static struct btrfs_block_group_cache *
 731 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 732 {
 733         return block_group_cache_tree_search(info, bytenr, 0);
 734 }
 735
 736 /*
 737  * return the block group that contains the given bytenr
 738  */
 739 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 740                                                  struct btrfs_fs_info *info,
 741                                                  u64 bytenr)
 742 {
 743         return block_group_cache_tree_search(info, bytenr, 1);
 744 }
 745
 746 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 747                                                   u64 flags)
 748 {
 749         struct list_head *head = &info->space_info;
 750         struct btrfs_space_info *found;
 751
 752         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 753
 754         rcu_read_lock();
 755         list_for_each_entry_rcu(found, head, list) {
 756                 if (found->flags & flags) {
 757                         rcu_read_unlock();
 758                         return found;
 759                 }
 760         }
 761         rcu_read_unlock();
 762         return NULL;
 763 }
 764
 765 /*
 766  * after adding space to the filesystem, we need to clear the full flags
 767  * on all the space infos.
 768  */
 769 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 770 {
 771         struct list_head *head = &info->space_info;
 772         struct btrfs_space_info *found;
 773
 774         rcu_read_lock();
 775         list_for_each_entry_rcu(found, head, list)
 776                 found->full = 0;
 777         rcu_read_unlock();
 778 }
 779
 780 /* simple helper to search for an existing data extent at a given offset */
 781 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 782 {
 783         int ret;
 784         struct btrfs_key key;
 785         struct btrfs_path *path;
 786
 787         path = btrfs_alloc_path();
 788         if (!path)
 789                 return -ENOMEM;
 790
 791         key.objectid = start;
 792         key.offset = len;
 793         key.type = BTRFS_EXTENT_ITEM_KEY;
 794         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 795                                 0, 0);
 796         btrfs_free_path(path);
 797         return ret;
 798 }
 799
 800 /*
 801  * helper function to lookup reference count and flags of a tree block.
 802  *
 803  * the head node for delayed ref is used to store the sum of all the
 804  * reference count modifications queued up in the rbtree. the head
 805  * node may also store the extent flags to set. This way you can check
 806  * to see what the reference count and extent flags would be if all of
 807  * the delayed refs are not processed.
 808  */
 809 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 810                              struct btrfs_root *root, u64 bytenr,
 811                              u64 offset, int metadata, u64 *refs, u64 *flags)
 812 {
 813         struct btrfs_delayed_ref_head *head;
 814         struct btrfs_delayed_ref_root *delayed_refs;
 815         struct btrfs_path *path;
 816         struct btrfs_extent_item *ei;
 817         struct extent_buffer *leaf;
 818         struct btrfs_key key;
 819         u32 item_size;
 820         u64 num_refs;
 821         u64 extent_flags;
 822         int ret;
 823
 824         /*
 825          * If we don't have skinny metadata, don't bother doing anything
 826          * different
 827          */
 828         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 829                 offset = root->nodesize;
 830                 metadata = 0;
 831         }
 832
 833         path = btrfs_alloc_path();
 834         if (!path)
 835                 return -ENOMEM;
 836
 837         if (!trans) {
 838                 path->skip_locking = 1;
 839                 path->search_commit_root = 1;
 840         }
 841
 842 search_again:
 843         key.objectid = bytenr;
 844         key.offset = offset;
 845         if (metadata)
 846                 key.type = BTRFS_METADATA_ITEM_KEY;
 847         else
 848                 key.type = BTRFS_EXTENT_ITEM_KEY;
 849
 850         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 851                                 &key, path, 0, 0);
 852         if (ret < 0)
 853                 goto out_free;
 854
 855         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 856                 if (path->slots[0]) {
 857                         path->slots[0]--;
 858                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 859                                               path->slots[0]);
 860                         if (key.objectid == bytenr &&
 861                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 862                             key.offset == root->nodesize)
 863                                 ret = 0;
 864                 }
 865         }
 866
 867         if (ret == 0) {
 868                 leaf = path->nodes[0];
 869                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 870                 if (item_size >= sizeof(*ei)) {
 871                         ei = btrfs_item_ptr(leaf, path->slots[0],
 872                                             struct btrfs_extent_item);
 873                         num_refs = btrfs_extent_refs(leaf, ei);
 874                         extent_flags = btrfs_extent_flags(leaf, ei);
 875                 } else {
 876 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 877                         struct btrfs_extent_item_v0 *ei0;
 878                         BUG_ON(item_size != sizeof(*ei0));
 879                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 880                                              struct btrfs_extent_item_v0);
 881                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 882                         /* FIXME: this isn't correct for data */
 883                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 884 #else
 885                         BUG();
 886 #endif
 887                 }
 888                 BUG_ON(num_refs == 0);
 889         } else {
 890                 num_refs = 0;
 891                 extent_flags = 0;
 892                 ret = 0;
 893         }
 894
 895         if (!trans)
 896                 goto out;
 897
 898         delayed_refs = &trans->transaction->delayed_refs;
 899         spin_lock(&delayed_refs->lock);
 900         head = btrfs_find_delayed_ref_head(trans, bytenr);
 901         if (head) {
 902                 if (!mutex_trylock(&head->mutex)) {
 903                         atomic_inc(&head->node.refs);
 904                         spin_unlock(&delayed_refs->lock);
 905
 906                         btrfs_release_path(path);
 907
 908                         /*
 909                          * Mutex was contended, block until it's released and try
 910                          * again
 911                          */
 912                         mutex_lock(&head->mutex);
 913                         mutex_unlock(&head->mutex);
 914                         btrfs_put_delayed_ref(&head->node);
 915                         goto search_again;
 916                 }
 917                 spin_lock(&head->lock);
 918                 if (head->extent_op && head->extent_op->update_flags)
 919                         extent_flags |= head->extent_op->flags_to_set;
 920                 else
 921                         BUG_ON(num_refs == 0);
 922
 923                 num_refs += head->node.ref_mod;
 924                 spin_unlock(&head->lock);
 925                 mutex_unlock(&head->mutex);
 926         }
 927         spin_unlock(&delayed_refs->lock);
 928 out:
 929         WARN_ON(num_refs == 0);
 930         if (refs)
 931                 *refs = num_refs;
 932         if (flags)
 933                 *flags = extent_flags;
 934 out_free:
 935         btrfs_free_path(path);
 936         return ret;
 937 }
 938
 939 /*
 940  * Back reference rules.  Back refs have three main goals:
 941  *
 942  * 1) differentiate between all holders of references to an extent so that
 943  *    when a reference is dropped we can make sure it was a valid reference
 944  *    before freeing the extent.
 945  *
 946  * 2) Provide enough information to quickly find the holders of an extent
 947  *    if we notice a given block is corrupted or bad.
 948  *
 949  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 950  *    maintenance.  This is actually the same as #2, but with a slightly
 951  *    different use case.
 952  *
 953  * There are two kinds of back refs. The implicit back refs is optimized
 954  * for pointers in non-shared tree blocks. For a given pointer in a block,
 955  * back refs of this kind provide information about the block's owner tree
 956  * and the pointer's key. These information allow us to find the block by
 957  * b-tree searching. The full back refs is for pointers in tree blocks not
 958  * referenced by their owner trees. The location of tree block is recorded
 959  * in the back refs. Actually the full back refs is generic, and can be
 960  * used in all cases the implicit back refs is used. The major shortcoming
 961  * of the full back refs is its overhead. Every time a tree block gets
 962  * COWed, we have to update back refs entry for all pointers in it.
 963  *
 964  * For a newly allocated tree block, we use implicit back refs for
 965  * pointers in it. This means most tree related operations only involve
 966  * implicit back refs. For a tree block created in old transaction, the
 967  * only way to drop a reference to it is COW it. So we can detect the
 968  * event that tree block loses its owner tree's reference and do the
 969  * back refs conversion.
 970  *
 971  * When a tree block is COWed through a tree, there are four cases:
 972  *
 973  * The reference count of the block is one and the tree is the block's
 974  * owner tree. Nothing to do in this case.
 975  *
 976  * The reference count of the block is one and the tree is not the
 977  * block's owner tree. In this case, full back refs is used for pointers
 978  * in the block. Remove these full back refs, add implicit back refs for
 979  * every pointers in the new block.
 980  *
 981  * The reference count of the block is greater than one and the tree is
 982  * the block's owner tree. In this case, implicit back refs is used for
 983  * pointers in the block. Add full back refs for every pointers in the
 984  * block, increase lower level extents' reference counts. The original
 985  * implicit back refs are entailed to the new block.
 986  *
 987  * The reference count of the block is greater than one and the tree is
 988  * not the block's owner tree. Add implicit back refs for every pointer in
 989  * the new block, increase lower level extents' reference count.
 990  *
 991  * Back Reference Key composing:
 992  *
 993  * The key objectid corresponds to the first byte in the extent,
 994  * The key type is used to differentiate between types of back refs.
 995  * There are different meanings of the key offset for different types
 996  * of back refs.
 997  *
 998  * File extents can be referenced by:
 999  *
1000  * - multiple snapshots, subvolumes, or different generations in one subvol
1001  * - different files inside a single subvolume
1002  * - different offsets inside a file (bookend extents in file.c)
1003  *
1004  * The extent ref structure for the implicit back refs has fields for:
1005  *
1006  * - Objectid of the subvolume root
1007  * - objectid of the file holding the reference
1008  * - original offset in the file
1009  * - how many bookend extents
1010  *
1011  * The key offset for the implicit back refs is hash of the first
1012  * three fields.
1013  *
1014  * The extent ref structure for the full back refs has field for:
1015  *
1016  * - number of pointers in the tree leaf
1017  *
1018  * The key offset for the implicit back refs is the first byte of
1019  * the tree leaf
1020  *
1021  * When a file extent is allocated, The implicit back refs is used.
1022  * the fields are filled in:
1023  *
1024  *     (root_key.objectid, inode objectid, offset in file, 1)
1025  *
1026  * When a file extent is removed file truncation, we find the
1027  * corresponding implicit back refs and check the following fields:
1028  *
1029  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1030  *
1031  * Btree extents can be referenced by:
1032  *
1033  * - Different subvolumes
1034  *
1035  * Both the implicit back refs and the full back refs for tree blocks
1036  * only consist of key. The key offset for the implicit back refs is
1037  * objectid of block's owner tree. The key offset for the full back refs
1038  * is the first byte of parent block.
1039  *
1040  * When implicit back refs is used, information about the lowest key and
1041  * level of the tree block are required. These information are stored in
1042  * tree block info structure.
1043  */
1044
1045 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1046 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1047                                   struct btrfs_root *root,
1048                                   struct btrfs_path *path,
1049                                   u64 owner, u32 extra_size)
1050 {
1051         struct btrfs_extent_item *item;
1052         struct btrfs_extent_item_v0 *ei0;
1053         struct btrfs_extent_ref_v0 *ref0;
1054         struct btrfs_tree_block_info *bi;
1055         struct extent_buffer *leaf;
1056         struct btrfs_key key;
1057         struct btrfs_key found_key;
1058         u32 new_size = sizeof(*item);
1059         u64 refs;
1060         int ret;
1061
1062         leaf = path->nodes[0];
1063         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1064
1065         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1066         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1067                              struct btrfs_extent_item_v0);
1068         refs = btrfs_extent_refs_v0(leaf, ei0);
1069
1070         if (owner == (u64)-1) {
1071                 while (1) {
1072                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1073                                 ret = btrfs_next_leaf(root, path);
1074                                 if (ret < 0)
1075                                         return ret;
1076                                 BUG_ON(ret > 0); /* Corruption */
1077                                 leaf = path->nodes[0];
1078                         }
1079                         btrfs_item_key_to_cpu(leaf, &found_key,
1080                                               path->slots[0]);
1081                         BUG_ON(key.objectid != found_key.objectid);
1082                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1083                                 path->slots[0]++;
1084                                 continue;
1085                         }
1086                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1087                                               struct btrfs_extent_ref_v0);
1088                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1089                         break;
1090                 }
1091         }
1092         btrfs_release_path(path);
1093
1094         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1095                 new_size += sizeof(*bi);
1096
1097         new_size -= sizeof(*ei0);
1098         ret = btrfs_search_slot(trans, root, &key, path,
1099                                 new_size + extra_size, 1);
1100         if (ret < 0)
1101                 return ret;
1102         BUG_ON(ret); /* Corruption */
1103
1104         btrfs_extend_item(root, path, new_size);
1105
1106         leaf = path->nodes[0];
1107         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1108         btrfs_set_extent_refs(leaf, item, refs);
1109         /* FIXME: get real generation */
1110         btrfs_set_extent_generation(leaf, item, 0);
1111         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1112                 btrfs_set_extent_flags(leaf, item,
1113                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1114                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1115                 bi = (struct btrfs_tree_block_info *)(item + 1);
1116                 /* FIXME: get first key of the block */
1117                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1118                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1119         } else {
1120                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1121         }
1122         btrfs_mark_buffer_dirty(leaf);
1123         return 0;
1124 }
1125 #endif
1126
1127 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1128 {
1129         u32 high_crc = ~(u32)0;
1130         u32 low_crc = ~(u32)0;
1131         __le64 lenum;
1132
1133         lenum = cpu_to_le64(root_objectid);
1134         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1135         lenum = cpu_to_le64(owner);
1136         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1137         lenum = cpu_to_le64(offset);
1138         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1139
1140         return ((u64)high_crc << 31) ^ (u64)low_crc;
1141 }
1142
1143 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1144                                      struct btrfs_extent_data_ref *ref)
1145 {
1146         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1147                                     btrfs_extent_data_ref_objectid(leaf, ref),
1148                                     btrfs_extent_data_ref_offset(leaf, ref));
1149 }
1150
1151 static int match_extent_data_ref(struct extent_buffer *leaf,
1152                                  struct btrfs_extent_data_ref *ref,
1153                                  u64 root_objectid, u64 owner, u64 offset)
1154 {
1155         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1156             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1157             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1158                 return 0;
1159         return 1;
1160 }
1161
1162 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1163                                            struct btrfs_root *root,
1164                                            struct btrfs_path *path,
1165                                            u64 bytenr, u64 parent,
1166                                            u64 root_objectid,
1167                                            u64 owner, u64 offset)
1168 {
1169         struct btrfs_key key;
1170         struct btrfs_extent_data_ref *ref;
1171         struct extent_buffer *leaf;
1172         u32 nritems;
1173         int ret;
1174         int recow;
1175         int err = -ENOENT;
1176
1177         key.objectid = bytenr;
1178         if (parent) {
1179                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1180                 key.offset = parent;
1181         } else {
1182                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1183                 key.offset = hash_extent_data_ref(root_objectid,
1184                                                   owner, offset);
1185         }
1186 again:
1187         recow = 0;
1188         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1189         if (ret < 0) {
1190                 err = ret;
1191                 goto fail;
1192         }
1193
1194         if (parent) {
1195                 if (!ret)
1196                         return 0;
1197 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1198                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1199                 btrfs_release_path(path);
1200                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1201                 if (ret < 0) {
1202                         err = ret;
1203                         goto fail;
1204                 }
1205                 if (!ret)
1206                         return 0;
1207 #endif
1208                 goto fail;
1209         }
1210
1211         leaf = path->nodes[0];
1212         nritems = btrfs_header_nritems(leaf);
1213         while (1) {
1214                 if (path->slots[0] >= nritems) {
1215                         ret = btrfs_next_leaf(root, path);
1216                         if (ret < 0)
1217                                 err = ret;
1218                         if (ret)
1219                                 goto fail;
1220
1221                         leaf = path->nodes[0];
1222                         nritems = btrfs_header_nritems(leaf);
1223                         recow = 1;
1224                 }
1225
1226                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1227                 if (key.objectid != bytenr ||
1228                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1229                         goto fail;
1230
1231                 ref = btrfs_item_ptr(leaf, path->slots[0],
1232                                      struct btrfs_extent_data_ref);
1233
1234                 if (match_extent_data_ref(leaf, ref, root_objectid,
1235                                           owner, offset)) {
1236                         if (recow) {
1237                                 btrfs_release_path(path);
1238                                 goto again;
1239                         }
1240                         err = 0;
1241                         break;
1242                 }
1243                 path->slots[0]++;
1244         }
1245 fail:
1246         return err;
1247 }
1248
1249 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1250                                            struct btrfs_root *root,
1251                                            struct btrfs_path *path,
1252                                            u64 bytenr, u64 parent,
1253                                            u64 root_objectid, u64 owner,
1254                                            u64 offset, int refs_to_add)
1255 {
1256         struct btrfs_key key;
1257         struct extent_buffer *leaf;
1258         u32 size;
1259         u32 num_refs;
1260         int ret;
1261
1262         key.objectid = bytenr;
1263         if (parent) {
1264                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1265                 key.offset = parent;
1266                 size = sizeof(struct btrfs_shared_data_ref);
1267         } else {
1268                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1269                 key.offset = hash_extent_data_ref(root_objectid,
1270                                                   owner, offset);
1271                 size = sizeof(struct btrfs_extent_data_ref);
1272         }
1273
1274         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1275         if (ret && ret != -EEXIST)
1276                 goto fail;
1277
1278         leaf = path->nodes[0];
1279         if (parent) {
1280                 struct btrfs_shared_data_ref *ref;
1281                 ref = btrfs_item_ptr(leaf, path->slots[0],
1282                                      struct btrfs_shared_data_ref);
1283                 if (ret == 0) {
1284                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1285                 } else {
1286                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1287                         num_refs += refs_to_add;
1288                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1289                 }
1290         } else {
1291                 struct btrfs_extent_data_ref *ref;
1292                 while (ret == -EEXIST) {
1293                         ref = btrfs_item_ptr(leaf, path->slots[0],
1294                                              struct btrfs_extent_data_ref);
1295                         if (match_extent_data_ref(leaf, ref, root_objectid,
1296                                                   owner, offset))
1297                                 break;
1298                         btrfs_release_path(path);
1299                         key.offset++;
1300                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1301                                                       size);
1302                         if (ret && ret != -EEXIST)
1303                                 goto fail;
1304
1305                         leaf = path->nodes[0];
1306                 }
1307                 ref = btrfs_item_ptr(leaf, path->slots[0],
1308                                      struct btrfs_extent_data_ref);
1309                 if (ret == 0) {
1310                         btrfs_set_extent_data_ref_root(leaf, ref,
1311                                                        root_objectid);
1312                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1313                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1314                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1315                 } else {
1316                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1317                         num_refs += refs_to_add;
1318                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1319                 }
1320         }
1321         btrfs_mark_buffer_dirty(leaf);
1322         ret = 0;
1323 fail:
1324         btrfs_release_path(path);
1325         return ret;
1326 }
1327
1328 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1329                                            struct btrfs_root *root,
1330                                            struct btrfs_path *path,
1331                                            int refs_to_drop, int *last_ref)
1332 {
1333         struct btrfs_key key;
1334         struct btrfs_extent_data_ref *ref1 = NULL;
1335         struct btrfs_shared_data_ref *ref2 = NULL;
1336         struct extent_buffer *leaf;
1337         u32 num_refs = 0;
1338         int ret = 0;
1339
1340         leaf = path->nodes[0];
1341         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1342
1343         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1344                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1345                                       struct btrfs_extent_data_ref);
1346                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1347         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1348                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1349                                       struct btrfs_shared_data_ref);
1350                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1351 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1352         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1353                 struct btrfs_extent_ref_v0 *ref0;
1354                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1355                                       struct btrfs_extent_ref_v0);
1356                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1357 #endif
1358         } else {
1359                 BUG();
1360         }
1361
1362         BUG_ON(num_refs < refs_to_drop);
1363         num_refs -= refs_to_drop;
1364
1365         if (num_refs == 0) {
1366                 ret = btrfs_del_item(trans, root, path);
1367                 *last_ref = 1;
1368         } else {
1369                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1370                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1371                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1372                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1373 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1374                 else {
1375                         struct btrfs_extent_ref_v0 *ref0;
1376                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1377                                         struct btrfs_extent_ref_v0);
1378                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1379                 }
1380 #endif
1381                 btrfs_mark_buffer_dirty(leaf);
1382         }
1383         return ret;
1384 }
1385
1386 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1387                                           struct btrfs_extent_inline_ref *iref)
1388 {
1389         struct btrfs_key key;
1390         struct extent_buffer *leaf;
1391         struct btrfs_extent_data_ref *ref1;
1392         struct btrfs_shared_data_ref *ref2;
1393         u32 num_refs = 0;
1394
1395         leaf = path->nodes[0];
1396         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1397         if (iref) {
1398                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1399                     BTRFS_EXTENT_DATA_REF_KEY) {
1400                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1401                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1402                 } else {
1403                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1404                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1405                 }
1406         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1407                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1408                                       struct btrfs_extent_data_ref);
1409                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1410         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1411                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1412                                       struct btrfs_shared_data_ref);
1413                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1414 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1415         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1416                 struct btrfs_extent_ref_v0 *ref0;
1417                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1418                                       struct btrfs_extent_ref_v0);
1419                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1420 #endif
1421         } else {
1422                 WARN_ON(1);
1423         }
1424         return num_refs;
1425 }
1426
1427 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1428                                           struct btrfs_root *root,
1429                                           struct btrfs_path *path,
1430                                           u64 bytenr, u64 parent,
1431                                           u64 root_objectid)
1432 {
1433         struct btrfs_key key;
1434         int ret;
1435
1436         key.objectid = bytenr;
1437         if (parent) {
1438                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1439                 key.offset = parent;
1440         } else {
1441                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1442                 key.offset = root_objectid;
1443         }
1444
1445         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1446         if (ret > 0)
1447                 ret = -ENOENT;
1448 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1449         if (ret == -ENOENT && parent) {
1450                 btrfs_release_path(path);
1451                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1452                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1453                 if (ret > 0)
1454                         ret = -ENOENT;
1455         }
1456 #endif
1457         return ret;
1458 }
1459
1460 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1461                                           struct btrfs_root *root,
1462                                           struct btrfs_path *path,
1463                                           u64 bytenr, u64 parent,
1464                                           u64 root_objectid)
1465 {
1466         struct btrfs_key key;
1467         int ret;
1468
1469         key.objectid = bytenr;
1470         if (parent) {
1471                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1472                 key.offset = parent;
1473         } else {
1474                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1475                 key.offset = root_objectid;
1476         }
1477
1478         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1479         btrfs_release_path(path);
1480         return ret;
1481 }
1482
1483 static inline int extent_ref_type(u64 parent, u64 owner)
1484 {
1485         int type;
1486         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1487                 if (parent > 0)
1488                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1489                 else
1490                         type = BTRFS_TREE_BLOCK_REF_KEY;
1491         } else {
1492                 if (parent > 0)
1493                         type = BTRFS_SHARED_DATA_REF_KEY;
1494                 else
1495                         type = BTRFS_EXTENT_DATA_REF_KEY;
1496         }
1497         return type;
1498 }
1499
1500 static int find_next_key(struct btrfs_path *path, int level,
1501                          struct btrfs_key *key)
1502
1503 {
1504         for (; level < BTRFS_MAX_LEVEL; level++) {
1505                 if (!path->nodes[level])
1506                         break;
1507                 if (path->slots[level] + 1 >=
1508                     btrfs_header_nritems(path->nodes[level]))
1509                         continue;
1510                 if (level == 0)
1511                         btrfs_item_key_to_cpu(path->nodes[level], key,
1512                                               path->slots[level] + 1);
1513                 else
1514                         btrfs_node_key_to_cpu(path->nodes[level], key,
1515                                               path->slots[level] + 1);
1516                 return 0;
1517         }
1518         return 1;
1519 }
1520
1521 /*
1522  * look for inline back ref. if back ref is found, *ref_ret is set
1523  * to the address of inline back ref, and 0 is returned.
1524  *
1525  * if back ref isn't found, *ref_ret is set to the address where it
1526  * should be inserted, and -ENOENT is returned.
1527  *
1528  * if insert is true and there are too many inline back refs, the path
1529  * points to the extent item, and -EAGAIN is returned.
1530  *
1531  * NOTE: inline back refs are ordered in the same way that back ref
1532  *       items in the tree are ordered.
1533  */
1534 static noinline_for_stack
1535 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1536                                  struct btrfs_root *root,
1537                                  struct btrfs_path *path,
1538                                  struct btrfs_extent_inline_ref **ref_ret,
1539                                  u64 bytenr, u64 num_bytes,
1540                                  u64 parent, u64 root_objectid,
1541                                  u64 owner, u64 offset, int insert)
1542 {
1543         struct btrfs_key key;
1544         struct extent_buffer *leaf;
1545         struct btrfs_extent_item *ei;
1546         struct btrfs_extent_inline_ref *iref;
1547         u64 flags;
1548         u64 item_size;
1549         unsigned long ptr;
1550         unsigned long end;
1551         int extra_size;
1552         int type;
1553         int want;
1554         int ret;
1555         int err = 0;
1556         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1557                                                  SKINNY_METADATA);
1558
1559         key.objectid = bytenr;
1560         key.type = BTRFS_EXTENT_ITEM_KEY;
1561         key.offset = num_bytes;
1562
1563         want = extent_ref_type(parent, owner);
1564         if (insert) {
1565                 extra_size = btrfs_extent_inline_ref_size(want);
1566                 path->keep_locks = 1;
1567         } else
1568                 extra_size = -1;
1569
1570         /*
1571          * Owner is our parent level, so we can just add one to get the level
1572          * for the block we are interested in.
1573          */
1574         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1575                 key.type = BTRFS_METADATA_ITEM_KEY;
1576                 key.offset = owner;
1577         }
1578
1579 again:
1580         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1581         if (ret < 0) {
1582                 err = ret;
1583                 goto out;
1584         }
1585
1586         /*
1587          * We may be a newly converted file system which still has the old fat
1588          * extent entries for metadata, so try and see if we have one of those.
1589          */
1590         if (ret > 0 && skinny_metadata) {
1591                 skinny_metadata = false;
1592                 if (path->slots[0]) {
1593                         path->slots[0]--;
1594                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1595                                               path->slots[0]);
1596                         if (key.objectid == bytenr &&
1597                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1598                             key.offset == num_bytes)
1599                                 ret = 0;
1600                 }
1601                 if (ret) {
1602                         key.objectid = bytenr;
1603                         key.type = BTRFS_EXTENT_ITEM_KEY;
1604                         key.offset = num_bytes;
1605                         btrfs_release_path(path);
1606                         goto again;
1607                 }
1608         }
1609
1610         if (ret && !insert) {
1611                 err = -ENOENT;
1612                 goto out;
1613         } else if (WARN_ON(ret)) {
1614                 err = -EIO;
1615                 goto out;
1616         }
1617
1618         leaf = path->nodes[0];
1619         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1620 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1621         if (item_size < sizeof(*ei)) {
1622                 if (!insert) {
1623                         err = -ENOENT;
1624                         goto out;
1625                 }
1626                 ret = convert_extent_item_v0(trans, root, path, owner,
1627                                              extra_size);
1628                 if (ret < 0) {
1629                         err = ret;
1630                         goto out;
1631                 }
1632                 leaf = path->nodes[0];
1633                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1634         }
1635 #endif
1636         BUG_ON(item_size < sizeof(*ei));
1637
1638         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1639         flags = btrfs_extent_flags(leaf, ei);
1640
1641         ptr = (unsigned long)(ei + 1);
1642         end = (unsigned long)ei + item_size;
1643
1644         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1645                 ptr += sizeof(struct btrfs_tree_block_info);
1646                 BUG_ON(ptr > end);
1647         }
1648
1649         err = -ENOENT;
1650         while (1) {
1651                 if (ptr >= end) {
1652                         WARN_ON(ptr > end);
1653                         break;
1654                 }
1655                 iref = (struct btrfs_extent_inline_ref *)ptr;
1656                 type = btrfs_extent_inline_ref_type(leaf, iref);
1657                 if (want < type)
1658                         break;
1659                 if (want > type) {
1660                         ptr += btrfs_extent_inline_ref_size(type);
1661                         continue;
1662                 }
1663
1664                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1665                         struct btrfs_extent_data_ref *dref;
1666                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1667                         if (match_extent_data_ref(leaf, dref, root_objectid,
1668                                                   owner, offset)) {
1669                                 err = 0;
1670                                 break;
1671                         }
1672                         if (hash_extent_data_ref_item(leaf, dref) <
1673                             hash_extent_data_ref(root_objectid, owner, offset))
1674                                 break;
1675                 } else {
1676                         u64 ref_offset;
1677                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1678                         if (parent > 0) {
1679                                 if (parent == ref_offset) {
1680                                         err = 0;
1681                                         break;
1682                                 }
1683                                 if (ref_offset < parent)
1684                                         break;
1685                         } else {
1686                                 if (root_objectid == ref_offset) {
1687                                         err = 0;
1688                                         break;
1689                                 }
1690                                 if (ref_offset < root_objectid)
1691                                         break;
1692                         }
1693                 }
1694                 ptr += btrfs_extent_inline_ref_size(type);
1695         }
1696         if (err == -ENOENT && insert) {
1697                 if (item_size + extra_size >=
1698                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1699                         err = -EAGAIN;
1700                         goto out;
1701                 }
1702                 /*
1703                  * To add new inline back ref, we have to make sure
1704                  * there is no corresponding back ref item.
1705                  * For simplicity, we just do not add new inline back
1706                  * ref if there is any kind of item for this block
1707                  */
1708                 if (find_next_key(path, 0, &key) == 0 &&
1709                     key.objectid == bytenr &&
1710                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1711                         err = -EAGAIN;
1712                         goto out;
1713                 }
1714         }
1715         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1716 out:
1717         if (insert) {
1718                 path->keep_locks = 0;
1719                 btrfs_unlock_up_safe(path, 1);
1720         }
1721         return err;
1722 }
1723
1724 /*
1725  * helper to add new inline back ref
1726  */
1727 static noinline_for_stack
1728 void setup_inline_extent_backref(struct btrfs_root *root,
1729                                  struct btrfs_path *path,
1730                                  struct btrfs_extent_inline_ref *iref,
1731                                  u64 parent, u64 root_objectid,
1732                                  u64 owner, u64 offset, int refs_to_add,
1733                                  struct btrfs_delayed_extent_op *extent_op)
1734 {
1735         struct extent_buffer *leaf;
1736         struct btrfs_extent_item *ei;
1737         unsigned long ptr;
1738         unsigned long end;
1739         unsigned long item_offset;
1740         u64 refs;
1741         int size;
1742         int type;
1743
1744         leaf = path->nodes[0];
1745         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1746         item_offset = (unsigned long)iref - (unsigned long)ei;
1747
1748         type = extent_ref_type(parent, owner);
1749         size = btrfs_extent_inline_ref_size(type);
1750
1751         btrfs_extend_item(root, path, size);
1752
1753         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1754         refs = btrfs_extent_refs(leaf, ei);
1755         refs += refs_to_add;
1756         btrfs_set_extent_refs(leaf, ei, refs);
1757         if (extent_op)
1758                 __run_delayed_extent_op(extent_op, leaf, ei);
1759
1760         ptr = (unsigned long)ei + item_offset;
1761         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1762         if (ptr < end - size)
1763                 memmove_extent_buffer(leaf, ptr + size, ptr,
1764                                       end - size - ptr);
1765
1766         iref = (struct btrfs_extent_inline_ref *)ptr;
1767         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1768         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1769                 struct btrfs_extent_data_ref *dref;
1770                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1771                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1772                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1773                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1774                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1775         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1776                 struct btrfs_shared_data_ref *sref;
1777                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1778                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1779                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1780         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1781                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1782         } else {
1783                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1784         }
1785         btrfs_mark_buffer_dirty(leaf);
1786 }
1787
1788 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1789                                  struct btrfs_root *root,
1790                                  struct btrfs_path *path,
1791                                  struct btrfs_extent_inline_ref **ref_ret,
1792                                  u64 bytenr, u64 num_bytes, u64 parent,
1793                                  u64 root_objectid, u64 owner, u64 offset)
1794 {
1795         int ret;
1796
1797         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1798                                            bytenr, num_bytes, parent,
1799                                            root_objectid, owner, offset, 0);
1800         if (ret != -ENOENT)
1801                 return ret;
1802
1803         btrfs_release_path(path);
1804         *ref_ret = NULL;
1805
1806         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1807                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1808                                             root_objectid);
1809         } else {
1810                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1811                                              root_objectid, owner, offset);
1812         }
1813         return ret;
1814 }
1815
1816 /*
1817  * helper to update/remove inline back ref
1818  */
1819 static noinline_for_stack
1820 void update_inline_extent_backref(struct btrfs_root *root,
1821                                   struct btrfs_path *path,
1822                                   struct btrfs_extent_inline_ref *iref,
1823                                   int refs_to_mod,
1824                                   struct btrfs_delayed_extent_op *extent_op,
1825                                   int *last_ref)
1826 {
1827         struct extent_buffer *leaf;
1828         struct btrfs_extent_item *ei;
1829         struct btrfs_extent_data_ref *dref = NULL;
1830         struct btrfs_shared_data_ref *sref = NULL;
1831         unsigned long ptr;
1832         unsigned long end;
1833         u32 item_size;
1834         int size;
1835         int type;
1836         u64 refs;
1837
1838         leaf = path->nodes[0];
1839         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1840         refs = btrfs_extent_refs(leaf, ei);
1841         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1842         refs += refs_to_mod;
1843         btrfs_set_extent_refs(leaf, ei, refs);
1844         if (extent_op)
1845                 __run_delayed_extent_op(extent_op, leaf, ei);
1846
1847         type = btrfs_extent_inline_ref_type(leaf, iref);
1848
1849         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1850                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1851                 refs = btrfs_extent_data_ref_count(leaf, dref);
1852         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1853                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1854                 refs = btrfs_shared_data_ref_count(leaf, sref);
1855         } else {
1856                 refs = 1;
1857                 BUG_ON(refs_to_mod != -1);
1858         }
1859
1860         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1861         refs += refs_to_mod;
1862
1863         if (refs > 0) {
1864                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1865                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1866                 else
1867                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1868         } else {
1869                 *last_ref = 1;
1870                 size =  btrfs_extent_inline_ref_size(type);
1871                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1872                 ptr = (unsigned long)iref;
1873                 end = (unsigned long)ei + item_size;
1874                 if (ptr + size < end)
1875                         memmove_extent_buffer(leaf, ptr, ptr + size,
1876                                               end - ptr - size);
1877                 item_size -= size;
1878                 btrfs_truncate_item(root, path, item_size, 1);
1879         }
1880         btrfs_mark_buffer_dirty(leaf);
1881 }
1882
1883 static noinline_for_stack
1884 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1885                                  struct btrfs_root *root,
1886                                  struct btrfs_path *path,
1887                                  u64 bytenr, u64 num_bytes, u64 parent,
1888                                  u64 root_objectid, u64 owner,
1889                                  u64 offset, int refs_to_add,
1890                                  struct btrfs_delayed_extent_op *extent_op)
1891 {
1892         struct btrfs_extent_inline_ref *iref;
1893         int ret;
1894
1895         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1896                                            bytenr, num_bytes, parent,
1897                                            root_objectid, owner, offset, 1);
1898         if (ret == 0) {
1899                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1900                 update_inline_extent_backref(root, path, iref,
1901                                              refs_to_add, extent_op, NULL);
1902         } else if (ret == -ENOENT) {
1903                 setup_inline_extent_backref(root, path, iref, parent,
1904                                             root_objectid, owner, offset,
1905                                             refs_to_add, extent_op);
1906                 ret = 0;
1907         }
1908         return ret;
1909 }
1910
1911 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1912                                  struct btrfs_root *root,
1913                                  struct btrfs_path *path,
1914                                  u64 bytenr, u64 parent, u64 root_objectid,
1915                                  u64 owner, u64 offset, int refs_to_add)
1916 {
1917         int ret;
1918         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1919                 BUG_ON(refs_to_add != 1);
1920                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1921                                             parent, root_objectid);
1922         } else {
1923                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1924                                              parent, root_objectid,
1925                                              owner, offset, refs_to_add);
1926         }
1927         return ret;
1928 }
1929
1930 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1931                                  struct btrfs_root *root,
1932                                  struct btrfs_path *path,
1933                                  struct btrfs_extent_inline_ref *iref,
1934                                  int refs_to_drop, int is_data, int *last_ref)
1935 {
1936         int ret = 0;
1937
1938         BUG_ON(!is_data && refs_to_drop != 1);
1939         if (iref) {
1940                 update_inline_extent_backref(root, path, iref,
1941                                              -refs_to_drop, NULL, last_ref);
1942         } else if (is_data) {
1943                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1944                                              last_ref);
1945         } else {
1946                 *last_ref = 1;
1947                 ret = btrfs_del_item(trans, root, path);
1948         }
1949         return ret;
1950 }
1951
1952 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1953 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1954                                u64 *discarded_bytes)
1955 {
1956         int j, ret = 0;
1957         u64 bytes_left, end;
1958         u64 aligned_start = ALIGN(start, 1 << 9);
1959
1960         if (WARN_ON(start != aligned_start)) {
1961                 len -= aligned_start - start;
1962                 len = round_down(len, 1 << 9);
1963                 start = aligned_start;
1964         }
1965
1966         *discarded_bytes = 0;
1967
1968         if (!len)
1969                 return 0;
1970
1971         end = start + len;
1972         bytes_left = len;
1973
1974         /* Skip any superblocks on this device. */
1975         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1976                 u64 sb_start = btrfs_sb_offset(j);
1977                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1978                 u64 size = sb_start - start;
1979
1980                 if (!in_range(sb_start, start, bytes_left) &&
1981                     !in_range(sb_end, start, bytes_left) &&
1982                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1983                         continue;
1984
1985                 /*
1986                  * Superblock spans beginning of range.  Adjust start and
1987                  * try again.
1988                  */
1989                 if (sb_start <= start) {
1990                         start += sb_end - start;
1991                         if (start > end) {
1992                                 bytes_left = 0;
1993                                 break;
1994                         }
1995                         bytes_left = end - start;
1996                         continue;
1997                 }
1998
1999                 if (size) {
2000                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2001                                                    GFP_NOFS, 0);
2002                         if (!ret)
2003                                 *discarded_bytes += size;
2004                         else if (ret != -EOPNOTSUPP)
2005                                 return ret;
2006                 }
2007
2008                 start = sb_end;
2009                 if (start > end) {
2010                         bytes_left = 0;
2011                         break;
2012                 }
2013                 bytes_left = end - start;
2014         }
2015
2016         if (bytes_left) {
2017                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2018                                            GFP_NOFS, 0);
2019                 if (!ret)
2020                         *discarded_bytes += bytes_left;
2021         }
2022         return ret;
2023 }
2024
2025 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2026                          u64 num_bytes, u64 *actual_bytes)
2027 {
2028         int ret;
2029         u64 discarded_bytes = 0;
2030         struct btrfs_bio *bbio = NULL;
2031
2032
2033         /*
2034          * Avoid races with device replace and make sure our bbio has devices
2035          * associated to its stripes that don't go away while we are discarding.
2036          */
2037         btrfs_bio_counter_inc_blocked(root->fs_info);
2038         /* Tell the block device(s) that the sectors can be discarded */
2039         ret = btrfs_map_block(root->fs_info, REQ_OP_DISCARD,
2040                               bytenr, &num_bytes, &bbio, 0);
2041         /* Error condition is -ENOMEM */
2042         if (!ret) {
2043                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2044                 int i;
2045
2046
2047                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2048                         u64 bytes;
2049                         if (!stripe->dev->can_discard)
2050                                 continue;
2051
2052                         ret = btrfs_issue_discard(stripe->dev->bdev,
2053                                                   stripe->physical,
2054                                                   stripe->length,
2055                                                   &bytes);
2056                         if (!ret)
2057                                 discarded_bytes += bytes;
2058                         else if (ret != -EOPNOTSUPP)
2059                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2060
2061                         /*
2062                          * Just in case we get back EOPNOTSUPP for some reason,
2063                          * just ignore the return value so we don't screw up
2064                          * people calling discard_extent.
2065                          */
2066                         ret = 0;
2067                 }
2068                 btrfs_put_bbio(bbio);
2069         }
2070         btrfs_bio_counter_dec(root->fs_info);
2071
2072         if (actual_bytes)
2073                 *actual_bytes = discarded_bytes;
2074
2075
2076         if (ret == -EOPNOTSUPP)
2077                 ret = 0;
2078         return ret;
2079 }
2080
2081 /* Can return -ENOMEM */
2082 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2083                          struct btrfs_root *root,
2084                          u64 bytenr, u64 num_bytes, u64 parent,
2085                          u64 root_objectid, u64 owner, u64 offset)
2086 {
2087         int ret;
2088         struct btrfs_fs_info *fs_info = root->fs_info;
2089
2090         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2091                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2092
2093         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2094                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2095                                         num_bytes,
2096                                         parent, root_objectid, (int)owner,
2097                                         BTRFS_ADD_DELAYED_REF, NULL);
2098         } else {
2099                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2100                                         num_bytes, parent, root_objectid,
2101                                         owner, offset, 0,
2102                                         BTRFS_ADD_DELAYED_REF, NULL);
2103         }
2104         return ret;
2105 }
2106
2107 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2108                                   struct btrfs_root *root,
2109                                   struct btrfs_delayed_ref_node *node,
2110                                   u64 parent, u64 root_objectid,
2111                                   u64 owner, u64 offset, int refs_to_add,
2112                                   struct btrfs_delayed_extent_op *extent_op)
2113 {
2114         struct btrfs_fs_info *fs_info = root->fs_info;
2115         struct btrfs_path *path;
2116         struct extent_buffer *leaf;
2117         struct btrfs_extent_item *item;
2118         struct btrfs_key key;
2119         u64 bytenr = node->bytenr;
2120         u64 num_bytes = node->num_bytes;
2121         u64 refs;
2122         int ret;
2123
2124         path = btrfs_alloc_path();
2125         if (!path)
2126                 return -ENOMEM;
2127
2128         path->reada = READA_FORWARD;
2129         path->leave_spinning = 1;
2130         /* this will setup the path even if it fails to insert the back ref */
2131         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2132                                            bytenr, num_bytes, parent,
2133                                            root_objectid, owner, offset,
2134                                            refs_to_add, extent_op);
2135         if ((ret < 0 && ret != -EAGAIN) || !ret)
2136                 goto out;
2137
2138         /*
2139          * Ok we had -EAGAIN which means we didn't have space to insert and
2140          * inline extent ref, so just update the reference count and add a
2141          * normal backref.
2142          */
2143         leaf = path->nodes[0];
2144         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2145         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2146         refs = btrfs_extent_refs(leaf, item);
2147         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2148         if (extent_op)
2149                 __run_delayed_extent_op(extent_op, leaf, item);
2150
2151         btrfs_mark_buffer_dirty(leaf);
2152         btrfs_release_path(path);
2153
2154         path->reada = READA_FORWARD;
2155         path->leave_spinning = 1;
2156         /* now insert the actual backref */
2157         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2158                                     path, bytenr, parent, root_objectid,
2159                                     owner, offset, refs_to_add);
2160         if (ret)
2161                 btrfs_abort_transaction(trans, ret);
2162 out:
2163         btrfs_free_path(path);
2164         return ret;
2165 }
2166
2167 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2168                                 struct btrfs_root *root,
2169                                 struct btrfs_delayed_ref_node *node,
2170                                 struct btrfs_delayed_extent_op *extent_op,
2171                                 int insert_reserved)
2172 {
2173         int ret = 0;
2174         struct btrfs_delayed_data_ref *ref;
2175         struct btrfs_key ins;
2176         u64 parent = 0;
2177         u64 ref_root = 0;
2178         u64 flags = 0;
2179
2180         ins.objectid = node->bytenr;
2181         ins.offset = node->num_bytes;
2182         ins.type = BTRFS_EXTENT_ITEM_KEY;
2183
2184         ref = btrfs_delayed_node_to_data_ref(node);
2185         trace_run_delayed_data_ref(root->fs_info, node, ref, node->action);
2186
2187         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2188                 parent = ref->parent;
2189         ref_root = ref->root;
2190
2191         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2192                 if (extent_op)
2193                         flags |= extent_op->flags_to_set;
2194                 ret = alloc_reserved_file_extent(trans, root,
2195                                                  parent, ref_root, flags,
2196                                                  ref->objectid, ref->offset,
2197                                                  &ins, node->ref_mod);
2198         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2199                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2200                                              ref_root, ref->objectid,
2201                                              ref->offset, node->ref_mod,
2202                                              extent_op);
2203         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2204                 ret = __btrfs_free_extent(trans, root, node, parent,
2205                                           ref_root, ref->objectid,
2206                                           ref->offset, node->ref_mod,
2207                                           extent_op);
2208         } else {
2209                 BUG();
2210         }
2211         return ret;
2212 }
2213
2214 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2215                                     struct extent_buffer *leaf,
2216                                     struct btrfs_extent_item *ei)
2217 {
2218         u64 flags = btrfs_extent_flags(leaf, ei);
2219         if (extent_op->update_flags) {
2220                 flags |= extent_op->flags_to_set;
2221                 btrfs_set_extent_flags(leaf, ei, flags);
2222         }
2223
2224         if (extent_op->update_key) {
2225                 struct btrfs_tree_block_info *bi;
2226                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2227                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2228                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2229         }
2230 }
2231
2232 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2233                                  struct btrfs_root *root,
2234                                  struct btrfs_delayed_ref_node *node,
2235                                  struct btrfs_delayed_extent_op *extent_op)
2236 {
2237         struct btrfs_key key;
2238         struct btrfs_path *path;
2239         struct btrfs_extent_item *ei;
2240         struct extent_buffer *leaf;
2241         u32 item_size;
2242         int ret;
2243         int err = 0;
2244         int metadata = !extent_op->is_data;
2245
2246         if (trans->aborted)
2247                 return 0;
2248
2249         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2250                 metadata = 0;
2251
2252         path = btrfs_alloc_path();
2253         if (!path)
2254                 return -ENOMEM;
2255
2256         key.objectid = node->bytenr;
2257
2258         if (metadata) {
2259                 key.type = BTRFS_METADATA_ITEM_KEY;
2260                 key.offset = extent_op->level;
2261         } else {
2262                 key.type = BTRFS_EXTENT_ITEM_KEY;
2263                 key.offset = node->num_bytes;
2264         }
2265
2266 again:
2267         path->reada = READA_FORWARD;
2268         path->leave_spinning = 1;
2269         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2270                                 path, 0, 1);
2271         if (ret < 0) {
2272                 err = ret;
2273                 goto out;
2274         }
2275         if (ret > 0) {
2276                 if (metadata) {
2277                         if (path->slots[0] > 0) {
2278                                 path->slots[0]--;
2279                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2280                                                       path->slots[0]);
2281                                 if (key.objectid == node->bytenr &&
2282                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2283                                     key.offset == node->num_bytes)
2284                                         ret = 0;
2285                         }
2286                         if (ret > 0) {
2287                                 btrfs_release_path(path);
2288                                 metadata = 0;
2289
2290                                 key.objectid = node->bytenr;
2291                                 key.offset = node->num_bytes;
2292                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2293                                 goto again;
2294                         }
2295                 } else {
2296                         err = -EIO;
2297                         goto out;
2298                 }
2299         }
2300
2301         leaf = path->nodes[0];
2302         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2303 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2304         if (item_size < sizeof(*ei)) {
2305                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2306                                              path, (u64)-1, 0);
2307                 if (ret < 0) {
2308                         err = ret;
2309                         goto out;
2310                 }
2311                 leaf = path->nodes[0];
2312                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2313         }
2314 #endif
2315         BUG_ON(item_size < sizeof(*ei));
2316         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2317         __run_delayed_extent_op(extent_op, leaf, ei);
2318
2319         btrfs_mark_buffer_dirty(leaf);
2320 out:
2321         btrfs_free_path(path);
2322         return err;
2323 }
2324
2325 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2326                                 struct btrfs_root *root,
2327                                 struct btrfs_delayed_ref_node *node,
2328                                 struct btrfs_delayed_extent_op *extent_op,
2329                                 int insert_reserved)
2330 {
2331         int ret = 0;
2332         struct btrfs_delayed_tree_ref *ref;
2333         struct btrfs_key ins;
2334         u64 parent = 0;
2335         u64 ref_root = 0;
2336         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2337                                                  SKINNY_METADATA);
2338
2339         ref = btrfs_delayed_node_to_tree_ref(node);
2340         trace_run_delayed_tree_ref(root->fs_info, node, ref, node->action);
2341
2342         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2343                 parent = ref->parent;
2344         ref_root = ref->root;
2345
2346         ins.objectid = node->bytenr;
2347         if (skinny_metadata) {
2348                 ins.offset = ref->level;
2349                 ins.type = BTRFS_METADATA_ITEM_KEY;
2350         } else {
2351                 ins.offset = node->num_bytes;
2352                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2353         }
2354
2355         if (node->ref_mod != 1) {
2356                 btrfs_err(root->fs_info,
2357         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2358                           node->bytenr, node->ref_mod, node->action, ref_root,
2359                           parent);
2360                 return -EIO;
2361         }
2362         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2363                 BUG_ON(!extent_op || !extent_op->update_flags);
2364                 ret = alloc_reserved_tree_block(trans, root,
2365                                                 parent, ref_root,
2366                                                 extent_op->flags_to_set,
2367                                                 &extent_op->key,
2368                                                 ref->level, &ins);
2369         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2370                 ret = __btrfs_inc_extent_ref(trans, root, node,
2371                                              parent, ref_root,
2372                                              ref->level, 0, 1,
2373                                              extent_op);
2374         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2375                 ret = __btrfs_free_extent(trans, root, node,
2376                                           parent, ref_root,
2377                                           ref->level, 0, 1, extent_op);
2378         } else {
2379                 BUG();
2380         }
2381         return ret;
2382 }
2383
2384 /* helper function to actually process a single delayed ref entry */
2385 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2386                                struct btrfs_root *root,
2387                                struct btrfs_delayed_ref_node *node,
2388                                struct btrfs_delayed_extent_op *extent_op,
2389                                int insert_reserved)
2390 {
2391         int ret = 0;
2392
2393         if (trans->aborted) {
2394                 if (insert_reserved)
2395                         btrfs_pin_extent(root, node->bytenr,
2396                                          node->num_bytes, 1);
2397                 return 0;
2398         }
2399
2400         if (btrfs_delayed_ref_is_head(node)) {
2401                 struct btrfs_delayed_ref_head *head;
2402                 /*
2403                  * we've hit the end of the chain and we were supposed
2404                  * to insert this extent into the tree.  But, it got
2405                  * deleted before we ever needed to insert it, so all
2406                  * we have to do is clean up the accounting
2407                  */
2408                 BUG_ON(extent_op);
2409                 head = btrfs_delayed_node_to_head(node);
2410                 trace_run_delayed_ref_head(root->fs_info, node, head,
2411                                            node->action);
2412
2413                 if (insert_reserved) {
2414                         btrfs_pin_extent(root, node->bytenr,
2415                                          node->num_bytes, 1);
2416                         if (head->is_data) {
2417                                 ret = btrfs_del_csums(trans, root,
2418                                                       node->bytenr,
2419                                                       node->num_bytes);
2420                         }
2421                 }
2422
2423                 /* Also free its reserved qgroup space */
2424                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2425                                               head->qgroup_ref_root,
2426                                               head->qgroup_reserved);
2427                 return ret;
2428         }
2429
2430         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2431             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2432                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2433                                            insert_reserved);
2434         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2435                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2436                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2437                                            insert_reserved);
2438         else
2439                 BUG();
2440         return ret;
2441 }
2442
2443 static inline struct btrfs_delayed_ref_node *
2444 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2445 {
2446         struct btrfs_delayed_ref_node *ref;
2447
2448         if (list_empty(&head->ref_list))
2449                 return NULL;
2450
2451         /*
2452          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2453          * This is to prevent a ref count from going down to zero, which deletes
2454          * the extent item from the extent tree, when there still are references
2455          * to add, which would fail because they would not find the extent item.
2456          */
2457         list_for_each_entry(ref, &head->ref_list, list) {
2458                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2459                         return ref;
2460         }
2461
2462         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2463                           list);
2464 }
2465
2466 /*
2467  * Returns 0 on success or if called with an already aborted transaction.
2468  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2469  */
2470 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2471                                              struct btrfs_root *root,
2472                                              unsigned long nr)
2473 {
2474         struct btrfs_delayed_ref_root *delayed_refs;
2475         struct btrfs_delayed_ref_node *ref;
2476         struct btrfs_delayed_ref_head *locked_ref = NULL;
2477         struct btrfs_delayed_extent_op *extent_op;
2478         struct btrfs_fs_info *fs_info = root->fs_info;
2479         ktime_t start = ktime_get();
2480         int ret;
2481         unsigned long count = 0;
2482         unsigned long actual_count = 0;
2483         int must_insert_reserved = 0;
2484
2485         delayed_refs = &trans->transaction->delayed_refs;
2486         while (1) {
2487                 if (!locked_ref) {
2488                         if (count >= nr)
2489                                 break;
2490
2491                         spin_lock(&delayed_refs->lock);
2492                         locked_ref = btrfs_select_ref_head(trans);
2493                         if (!locked_ref) {
2494                                 spin_unlock(&delayed_refs->lock);
2495                                 break;
2496                         }
2497
2498                         /* grab the lock that says we are going to process
2499                          * all the refs for this head */
2500                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2501                         spin_unlock(&delayed_refs->lock);
2502                         /*
2503                          * we may have dropped the spin lock to get the head
2504                          * mutex lock, and that might have given someone else
2505                          * time to free the head.  If that's true, it has been
2506                          * removed from our list and we can move on.
2507                          */
2508                         if (ret == -EAGAIN) {
2509                                 locked_ref = NULL;
2510                                 count++;
2511                                 continue;
2512                         }
2513                 }
2514
2515                 /*
2516                  * We need to try and merge add/drops of the same ref since we
2517                  * can run into issues with relocate dropping the implicit ref
2518                  * and then it being added back again before the drop can
2519                  * finish.  If we merged anything we need to re-loop so we can
2520                  * get a good ref.
2521                  * Or we can get node references of the same type that weren't
2522                  * merged when created due to bumps in the tree mod seq, and
2523                  * we need to merge them to prevent adding an inline extent
2524                  * backref before dropping it (triggering a BUG_ON at
2525                  * insert_inline_extent_backref()).
2526                  */
2527                 spin_lock(&locked_ref->lock);
2528                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2529                                          locked_ref);
2530
2531                 /*
2532                  * locked_ref is the head node, so we have to go one
2533                  * node back for any delayed ref updates
2534                  */
2535                 ref = select_delayed_ref(locked_ref);
2536
2537                 if (ref && ref->seq &&
2538                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2539                         spin_unlock(&locked_ref->lock);
2540                         spin_lock(&delayed_refs->lock);
2541                         locked_ref->processing = 0;
2542                         delayed_refs->num_heads_ready++;
2543                         spin_unlock(&delayed_refs->lock);
2544                         btrfs_delayed_ref_unlock(locked_ref);
2545                         locked_ref = NULL;
2546                         cond_resched();
2547                         count++;
2548                         continue;
2549                 }
2550
2551                 /*
2552                  * record the must insert reserved flag before we
2553                  * drop the spin lock.
2554                  */
2555                 must_insert_reserved = locked_ref->must_insert_reserved;
2556                 locked_ref->must_insert_reserved = 0;
2557
2558                 extent_op = locked_ref->extent_op;
2559                 locked_ref->extent_op = NULL;
2560
2561                 if (!ref) {
2562
2563
2564                         /* All delayed refs have been processed, Go ahead
2565                          * and send the head node to run_one_delayed_ref,
2566                          * so that any accounting fixes can happen
2567                          */
2568                         ref = &locked_ref->node;
2569
2570                         if (extent_op && must_insert_reserved) {
2571                                 btrfs_free_delayed_extent_op(extent_op);
2572                                 extent_op = NULL;
2573                         }
2574
2575                         if (extent_op) {
2576                                 spin_unlock(&locked_ref->lock);
2577                                 ret = run_delayed_extent_op(trans, root,
2578                                                             ref, extent_op);
2579                                 btrfs_free_delayed_extent_op(extent_op);
2580
2581                                 if (ret) {
2582                                         /*
2583                                          * Need to reset must_insert_reserved if
2584                                          * there was an error so the abort stuff
2585                                          * can cleanup the reserved space
2586                                          * properly.
2587                                          */
2588                                         if (must_insert_reserved)
2589                                                 locked_ref->must_insert_reserved = 1;
2590                                         spin_lock(&delayed_refs->lock);
2591                                         locked_ref->processing = 0;
2592                                         delayed_refs->num_heads_ready++;
2593                                         spin_unlock(&delayed_refs->lock);
2594                                         btrfs_debug(fs_info,
2595                                                     "run_delayed_extent_op returned %d",
2596                                                     ret);
2597                                         btrfs_delayed_ref_unlock(locked_ref);
2598                                         return ret;
2599                                 }
2600                                 continue;
2601                         }
2602
2603                         /*
2604                          * Need to drop our head ref lock and re-acquire the
2605                          * delayed ref lock and then re-check to make sure
2606                          * nobody got added.
2607                          */
2608                         spin_unlock(&locked_ref->lock);
2609                         spin_lock(&delayed_refs->lock);
2610                         spin_lock(&locked_ref->lock);
2611                         if (!list_empty(&locked_ref->ref_list) ||
2612                             locked_ref->extent_op) {
2613                                 spin_unlock(&locked_ref->lock);
2614                                 spin_unlock(&delayed_refs->lock);
2615                                 continue;
2616                         }
2617                         ref->in_tree = 0;
2618                         delayed_refs->num_heads--;
2619                         rb_erase(&locked_ref->href_node,
2620                                  &delayed_refs->href_root);
2621                         spin_unlock(&delayed_refs->lock);
2622                 } else {
2623                         actual_count++;
2624                         ref->in_tree = 0;
2625                         list_del(&ref->list);
2626                 }
2627                 atomic_dec(&delayed_refs->num_entries);
2628
2629                 if (!btrfs_delayed_ref_is_head(ref)) {
2630                         /*
2631                          * when we play the delayed ref, also correct the
2632                          * ref_mod on head
2633                          */
2634                         switch (ref->action) {
2635                         case BTRFS_ADD_DELAYED_REF:
2636                         case BTRFS_ADD_DELAYED_EXTENT:
2637                                 locked_ref->node.ref_mod -= ref->ref_mod;
2638                                 break;
2639                         case BTRFS_DROP_DELAYED_REF:
2640                                 locked_ref->node.ref_mod += ref->ref_mod;
2641                                 break;
2642                         default:
2643                                 WARN_ON(1);
2644                         }
2645                 }
2646                 spin_unlock(&locked_ref->lock);
2647
2648                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2649                                           must_insert_reserved);
2650
2651                 btrfs_free_delayed_extent_op(extent_op);
2652                 if (ret) {
2653                         spin_lock(&delayed_refs->lock);
2654                         locked_ref->processing = 0;
2655                         delayed_refs->num_heads_ready++;
2656                         spin_unlock(&delayed_refs->lock);
2657                         btrfs_delayed_ref_unlock(locked_ref);
2658                         btrfs_put_delayed_ref(ref);
2659                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2660                                     ret);
2661                         return ret;
2662                 }
2663
2664                 /*
2665                  * If this node is a head, that means all the refs in this head
2666                  * have been dealt with, and we will pick the next head to deal
2667                  * with, so we must unlock the head and drop it from the cluster
2668                  * list before we release it.
2669                  */
2670                 if (btrfs_delayed_ref_is_head(ref)) {
2671                         if (locked_ref->is_data &&
2672                             locked_ref->total_ref_mod < 0) {
2673                                 spin_lock(&delayed_refs->lock);
2674                                 delayed_refs->pending_csums -= ref->num_bytes;
2675                                 spin_unlock(&delayed_refs->lock);
2676                         }
2677                         btrfs_delayed_ref_unlock(locked_ref);
2678                         locked_ref = NULL;
2679                 }
2680                 btrfs_put_delayed_ref(ref);
2681                 count++;
2682                 cond_resched();
2683         }
2684
2685         /*
2686          * We don't want to include ref heads since we can have empty ref heads
2687          * and those will drastically skew our runtime down since we just do
2688          * accounting, no actual extent tree updates.
2689          */
2690         if (actual_count > 0) {
2691                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2692                 u64 avg;
2693
2694                 /*
2695                  * We weigh the current average higher than our current runtime
2696                  * to avoid large swings in the average.
2697                  */
2698                 spin_lock(&delayed_refs->lock);
2699                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2700                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2701                 spin_unlock(&delayed_refs->lock);
2702         }
2703         return 0;
2704 }
2705
2706 #ifdef SCRAMBLE_DELAYED_REFS
2707 /*
2708  * Normally delayed refs get processed in ascending bytenr order. This
2709  * correlates in most cases to the order added. To expose dependencies on this
2710  * order, we start to process the tree in the middle instead of the beginning
2711  */
2712 static u64 find_middle(struct rb_root *root)
2713 {
2714         struct rb_node *n = root->rb_node;
2715         struct btrfs_delayed_ref_node *entry;
2716         int alt = 1;
2717         u64 middle;
2718         u64 first = 0, last = 0;
2719
2720         n = rb_first(root);
2721         if (n) {
2722                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2723                 first = entry->bytenr;
2724         }
2725         n = rb_last(root);
2726         if (n) {
2727                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2728                 last = entry->bytenr;
2729         }
2730         n = root->rb_node;
2731
2732         while (n) {
2733                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2734                 WARN_ON(!entry->in_tree);
2735
2736                 middle = entry->bytenr;
2737
2738                 if (alt)
2739                         n = n->rb_left;
2740                 else
2741                         n = n->rb_right;
2742
2743                 alt = 1 - alt;
2744         }
2745         return middle;
2746 }
2747 #endif
2748
2749 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2750 {
2751         u64 num_bytes;
2752
2753         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2754                              sizeof(struct btrfs_extent_inline_ref));
2755         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2756                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2757
2758         /*
2759          * We don't ever fill up leaves all the way so multiply by 2 just to be
2760          * closer to what we're really going to want to use.
2761          */
2762         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2763 }
2764
2765 /*
2766  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2767  * would require to store the csums for that many bytes.
2768  */
2769 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2770 {
2771         u64 csum_size;
2772         u64 num_csums_per_leaf;
2773         u64 num_csums;
2774
2775         csum_size = BTRFS_MAX_ITEM_SIZE(root);
2776         num_csums_per_leaf = div64_u64(csum_size,
2777                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2778         num_csums = div64_u64(csum_bytes, root->sectorsize);
2779         num_csums += num_csums_per_leaf - 1;
2780         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2781         return num_csums;
2782 }
2783
2784 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2785                                        struct btrfs_root *root)
2786 {
2787         struct btrfs_block_rsv *global_rsv;
2788         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2789         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2790         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2791         u64 num_bytes, num_dirty_bgs_bytes;
2792         int ret = 0;
2793
2794         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2795         num_heads = heads_to_leaves(root, num_heads);
2796         if (num_heads > 1)
2797                 num_bytes += (num_heads - 1) * root->nodesize;
2798         num_bytes <<= 1;
2799         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2800         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2801                                                              num_dirty_bgs);
2802         global_rsv = &root->fs_info->global_block_rsv;
2803
2804         /*
2805          * If we can't allocate any more chunks lets make sure we have _lots_ of
2806          * wiggle room since running delayed refs can create more delayed refs.
2807          */
2808         if (global_rsv->space_info->full) {
2809                 num_dirty_bgs_bytes <<= 1;
2810                 num_bytes <<= 1;
2811         }
2812
2813         spin_lock(&global_rsv->lock);
2814         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2815                 ret = 1;
2816         spin_unlock(&global_rsv->lock);
2817         return ret;
2818 }
2819
2820 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2821                                        struct btrfs_root *root)
2822 {
2823         struct btrfs_fs_info *fs_info = root->fs_info;
2824         u64 num_entries =
2825                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2826         u64 avg_runtime;
2827         u64 val;
2828
2829         smp_mb();
2830         avg_runtime = fs_info->avg_delayed_ref_runtime;
2831         val = num_entries * avg_runtime;
2832         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2833                 return 1;
2834         if (val >= NSEC_PER_SEC / 2)
2835                 return 2;
2836
2837         return btrfs_check_space_for_delayed_refs(trans, root);
2838 }
2839
2840 struct async_delayed_refs {
2841         struct btrfs_root *root;
2842         u64 transid;
2843         int count;
2844         int error;
2845         int sync;
2846         struct completion wait;
2847         struct btrfs_work work;
2848 };
2849
2850 static void delayed_ref_async_start(struct btrfs_work *work)
2851 {
2852         struct async_delayed_refs *async;
2853         struct btrfs_trans_handle *trans;
2854         int ret;
2855
2856         async = container_of(work, struct async_delayed_refs, work);
2857
2858         /* if the commit is already started, we don't need to wait here */
2859         if (btrfs_transaction_blocked(async->root->fs_info))
2860                 goto done;
2861
2862         trans = btrfs_join_transaction(async->root);
2863         if (IS_ERR(trans)) {
2864                 async->error = PTR_ERR(trans);
2865                 goto done;
2866         }
2867
2868         /*
2869          * trans->sync means that when we call end_transaction, we won't
2870          * wait on delayed refs
2871          */
2872         trans->sync = true;
2873
2874         /* Don't bother flushing if we got into a different transaction */
2875         if (trans->transid > async->transid)
2876                 goto end;
2877
2878         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2879         if (ret)
2880                 async->error = ret;
2881 end:
2882         ret = btrfs_end_transaction(trans, async->root);
2883         if (ret && !async->error)
2884                 async->error = ret;
2885 done:
2886         if (async->sync)
2887                 complete(&async->wait);
2888         else
2889                 kfree(async);
2890 }
2891
2892 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2893                                  unsigned long count, u64 transid, int wait)
2894 {
2895         struct async_delayed_refs *async;
2896         int ret;
2897
2898         async = kmalloc(sizeof(*async), GFP_NOFS);
2899         if (!async)
2900                 return -ENOMEM;
2901
2902         async->root = root->fs_info->tree_root;
2903         async->count = count;
2904         async->error = 0;
2905         async->transid = transid;
2906         if (wait)
2907                 async->sync = 1;
2908         else
2909                 async->sync = 0;
2910         init_completion(&async->wait);
2911
2912         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2913                         delayed_ref_async_start, NULL, NULL);
2914
2915         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2916
2917         if (wait) {
2918                 wait_for_completion(&async->wait);
2919                 ret = async->error;
2920                 kfree(async);
2921                 return ret;
2922         }
2923         return 0;
2924 }
2925
2926 /*
2927  * this starts processing the delayed reference count updates and
2928  * extent insertions we have queued up so far.  count can be
2929  * 0, which means to process everything in the tree at the start
2930  * of the run (but not newly added entries), or it can be some target
2931  * number you'd like to process.
2932  *
2933  * Returns 0 on success or if called with an aborted transaction
2934  * Returns <0 on error and aborts the transaction
2935  */
2936 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2937                            struct btrfs_root *root, unsigned long count)
2938 {
2939         struct rb_node *node;
2940         struct btrfs_delayed_ref_root *delayed_refs;
2941         struct btrfs_delayed_ref_head *head;
2942         int ret;
2943         int run_all = count == (unsigned long)-1;
2944         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2945
2946         /* We'll clean this up in btrfs_cleanup_transaction */
2947         if (trans->aborted)
2948                 return 0;
2949
2950         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &root->fs_info->flags))
2951                 return 0;
2952
2953         if (root == root->fs_info->extent_root)
2954                 root = root->fs_info->tree_root;
2955
2956         delayed_refs = &trans->transaction->delayed_refs;
2957         if (count == 0)
2958                 count = atomic_read(&delayed_refs->num_entries) * 2;
2959
2960 again:
2961 #ifdef SCRAMBLE_DELAYED_REFS
2962         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2963 #endif
2964         trans->can_flush_pending_bgs = false;
2965         ret = __btrfs_run_delayed_refs(trans, root, count);
2966         if (ret < 0) {
2967                 btrfs_abort_transaction(trans, ret);
2968                 return ret;
2969         }
2970
2971         if (run_all) {
2972                 if (!list_empty(&trans->new_bgs))
2973                         btrfs_create_pending_block_groups(trans, root);
2974
2975                 spin_lock(&delayed_refs->lock);
2976                 node = rb_first(&delayed_refs->href_root);
2977                 if (!node) {
2978                         spin_unlock(&delayed_refs->lock);
2979                         goto out;
2980                 }
2981
2982                 while (node) {
2983                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2984                                         href_node);
2985                         if (btrfs_delayed_ref_is_head(&head->node)) {
2986                                 struct btrfs_delayed_ref_node *ref;
2987
2988                                 ref = &head->node;
2989                                 atomic_inc(&ref->refs);
2990
2991                                 spin_unlock(&delayed_refs->lock);
2992                                 /*
2993                                  * Mutex was contended, block until it's
2994                                  * released and try again
2995                                  */
2996                                 mutex_lock(&head->mutex);
2997                                 mutex_unlock(&head->mutex);
2998
2999                                 btrfs_put_delayed_ref(ref);
3000                                 cond_resched();
3001                                 goto again;
3002                         } else {
3003                                 WARN_ON(1);
3004                         }
3005                         node = rb_next(node);
3006                 }
3007                 spin_unlock(&delayed_refs->lock);
3008                 cond_resched();
3009                 goto again;
3010         }
3011 out:
3012         assert_qgroups_uptodate(trans);
3013         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3014         return 0;
3015 }
3016
3017 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3018                                 struct btrfs_root *root,
3019                                 u64 bytenr, u64 num_bytes, u64 flags,
3020                                 int level, int is_data)
3021 {
3022         struct btrfs_delayed_extent_op *extent_op;
3023         int ret;
3024
3025         extent_op = btrfs_alloc_delayed_extent_op();
3026         if (!extent_op)
3027                 return -ENOMEM;
3028
3029         extent_op->flags_to_set = flags;
3030         extent_op->update_flags = true;
3031         extent_op->update_key = false;
3032         extent_op->is_data = is_data ? true : false;
3033         extent_op->level = level;
3034
3035         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3036                                           num_bytes, extent_op);
3037         if (ret)
3038                 btrfs_free_delayed_extent_op(extent_op);
3039         return ret;
3040 }
3041
3042 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3043                                       struct btrfs_root *root,
3044                                       struct btrfs_path *path,
3045                                       u64 objectid, u64 offset, u64 bytenr)
3046 {
3047         struct btrfs_delayed_ref_head *head;
3048         struct btrfs_delayed_ref_node *ref;
3049         struct btrfs_delayed_data_ref *data_ref;
3050         struct btrfs_delayed_ref_root *delayed_refs;
3051         int ret = 0;
3052
3053         delayed_refs = &trans->transaction->delayed_refs;
3054         spin_lock(&delayed_refs->lock);
3055         head = btrfs_find_delayed_ref_head(trans, bytenr);
3056         if (!head) {
3057                 spin_unlock(&delayed_refs->lock);
3058                 return 0;
3059         }
3060
3061         if (!mutex_trylock(&head->mutex)) {
3062                 atomic_inc(&head->node.refs);
3063                 spin_unlock(&delayed_refs->lock);
3064
3065                 btrfs_release_path(path);
3066
3067                 /*
3068                  * Mutex was contended, block until it's released and let
3069                  * caller try again
3070                  */
3071                 mutex_lock(&head->mutex);
3072                 mutex_unlock(&head->mutex);
3073                 btrfs_put_delayed_ref(&head->node);
3074                 return -EAGAIN;
3075         }
3076         spin_unlock(&delayed_refs->lock);
3077
3078         spin_lock(&head->lock);
3079         list_for_each_entry(ref, &head->ref_list, list) {
3080                 /* If it's a shared ref we know a cross reference exists */
3081                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3082                         ret = 1;
3083                         break;
3084                 }
3085
3086                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3087
3088                 /*
3089                  * If our ref doesn't match the one we're currently looking at
3090                  * then we have a cross reference.
3091                  */
3092                 if (data_ref->root != root->root_key.objectid ||
3093                     data_ref->objectid != objectid ||
3094                     data_ref->offset != offset) {
3095                         ret = 1;
3096                         break;
3097                 }
3098         }
3099         spin_unlock(&head->lock);
3100         mutex_unlock(&head->mutex);
3101         return ret;
3102 }
3103
3104 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3105                                         struct btrfs_root *root,
3106                                         struct btrfs_path *path,
3107                                         u64 objectid, u64 offset, u64 bytenr)
3108 {
3109         struct btrfs_root *extent_root = root->fs_info->extent_root;
3110         struct extent_buffer *leaf;
3111         struct btrfs_extent_data_ref *ref;
3112         struct btrfs_extent_inline_ref *iref;
3113         struct btrfs_extent_item *ei;
3114         struct btrfs_key key;
3115         u32 item_size;
3116         int ret;
3117
3118         key.objectid = bytenr;
3119         key.offset = (u64)-1;
3120         key.type = BTRFS_EXTENT_ITEM_KEY;
3121
3122         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3123         if (ret < 0)
3124                 goto out;
3125         BUG_ON(ret == 0); /* Corruption */
3126
3127         ret = -ENOENT;
3128         if (path->slots[0] == 0)
3129                 goto out;
3130
3131         path->slots[0]--;
3132         leaf = path->nodes[0];
3133         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3134
3135         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3136                 goto out;
3137
3138         ret = 1;
3139         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3140 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3141         if (item_size < sizeof(*ei)) {
3142                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3143                 goto out;
3144         }
3145 #endif
3146         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3147
3148         if (item_size != sizeof(*ei) +
3149             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3150                 goto out;
3151
3152         if (btrfs_extent_generation(leaf, ei) <=
3153             btrfs_root_last_snapshot(&root->root_item))
3154                 goto out;
3155
3156         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3157         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3158             BTRFS_EXTENT_DATA_REF_KEY)
3159                 goto out;
3160
3161         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3162         if (btrfs_extent_refs(leaf, ei) !=
3163             btrfs_extent_data_ref_count(leaf, ref) ||
3164             btrfs_extent_data_ref_root(leaf, ref) !=
3165             root->root_key.objectid ||
3166             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3167             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3168                 goto out;
3169
3170         ret = 0;
3171 out:
3172         return ret;
3173 }
3174
3175 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3176                           struct btrfs_root *root,
3177                           u64 objectid, u64 offset, u64 bytenr)
3178 {
3179         struct btrfs_path *path;
3180         int ret;
3181         int ret2;
3182
3183         path = btrfs_alloc_path();
3184         if (!path)
3185                 return -ENOENT;
3186
3187         do {
3188                 ret = check_committed_ref(trans, root, path, objectid,
3189                                           offset, bytenr);
3190                 if (ret && ret != -ENOENT)
3191                         goto out;
3192
3193                 ret2 = check_delayed_ref(trans, root, path, objectid,
3194                                          offset, bytenr);
3195         } while (ret2 == -EAGAIN);
3196
3197         if (ret2 && ret2 != -ENOENT) {
3198                 ret = ret2;
3199                 goto out;
3200         }
3201
3202         if (ret != -ENOENT || ret2 != -ENOENT)
3203                 ret = 0;
3204 out:
3205         btrfs_free_path(path);
3206         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3207                 WARN_ON(ret > 0);
3208         return ret;
3209 }
3210
3211 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3212                            struct btrfs_root *root,
3213                            struct extent_buffer *buf,
3214                            int full_backref, int inc)
3215 {
3216         u64 bytenr;
3217         u64 num_bytes;
3218         u64 parent;
3219         u64 ref_root;
3220         u32 nritems;
3221         struct btrfs_key key;
3222         struct btrfs_file_extent_item *fi;
3223         int i;
3224         int level;
3225         int ret = 0;
3226         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3227                             u64, u64, u64, u64, u64, u64);
3228
3229
3230         if (btrfs_is_testing(root->fs_info))
3231                 return 0;
3232
3233         ref_root = btrfs_header_owner(buf);
3234         nritems = btrfs_header_nritems(buf);
3235         level = btrfs_header_level(buf);
3236
3237         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3238                 return 0;
3239
3240         if (inc)
3241                 process_func = btrfs_inc_extent_ref;
3242         else
3243                 process_func = btrfs_free_extent;
3244
3245         if (full_backref)
3246                 parent = buf->start;
3247         else
3248                 parent = 0;
3249
3250         for (i = 0; i < nritems; i++) {
3251                 if (level == 0) {
3252                         btrfs_item_key_to_cpu(buf, &key, i);
3253                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3254                                 continue;
3255                         fi = btrfs_item_ptr(buf, i,
3256                                             struct btrfs_file_extent_item);
3257                         if (btrfs_file_extent_type(buf, fi) ==
3258                             BTRFS_FILE_EXTENT_INLINE)
3259                                 continue;
3260                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3261                         if (bytenr == 0)
3262                                 continue;
3263
3264                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3265                         key.offset -= btrfs_file_extent_offset(buf, fi);
3266                         ret = process_func(trans, root, bytenr, num_bytes,
3267                                            parent, ref_root, key.objectid,
3268                                            key.offset);
3269                         if (ret)
3270                                 goto fail;
3271                 } else {
3272                         bytenr = btrfs_node_blockptr(buf, i);
3273                         num_bytes = root->nodesize;
3274                         ret = process_func(trans, root, bytenr, num_bytes,
3275                                            parent, ref_root, level - 1, 0);
3276                         if (ret)
3277                                 goto fail;
3278                 }
3279         }
3280         return 0;
3281 fail:
3282         return ret;
3283 }
3284
3285 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3286                   struct extent_buffer *buf, int full_backref)
3287 {
3288         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3289 }
3290
3291 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3292                   struct extent_buffer *buf, int full_backref)
3293 {
3294         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3295 }
3296
3297 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3298                                  struct btrfs_root *root,
3299                                  struct btrfs_path *path,
3300                                  struct btrfs_block_group_cache *cache)
3301 {
3302         int ret;
3303         struct btrfs_root *extent_root = root->fs_info->extent_root;
3304         unsigned long bi;
3305         struct extent_buffer *leaf;
3306
3307         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3308         if (ret) {
3309                 if (ret > 0)
3310                         ret = -ENOENT;
3311                 goto fail;
3312         }
3313
3314         leaf = path->nodes[0];
3315         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3316         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3317         btrfs_mark_buffer_dirty(leaf);
3318 fail:
3319         btrfs_release_path(path);
3320         return ret;
3321
3322 }
3323
3324 static struct btrfs_block_group_cache *
3325 next_block_group(struct btrfs_root *root,
3326                  struct btrfs_block_group_cache *cache)
3327 {
3328         struct rb_node *node;
3329
3330         spin_lock(&root->fs_info->block_group_cache_lock);
3331
3332         /* If our block group was removed, we need a full search. */
3333         if (RB_EMPTY_NODE(&cache->cache_node)) {
3334                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3335
3336                 spin_unlock(&root->fs_info->block_group_cache_lock);
3337                 btrfs_put_block_group(cache);
3338                 cache = btrfs_lookup_first_block_group(root->fs_info,
3339                                                        next_bytenr);
3340                 return cache;
3341         }
3342         node = rb_next(&cache->cache_node);
3343         btrfs_put_block_group(cache);
3344         if (node) {
3345                 cache = rb_entry(node, struct btrfs_block_group_cache,
3346                                  cache_node);
3347                 btrfs_get_block_group(cache);
3348         } else
3349                 cache = NULL;
3350         spin_unlock(&root->fs_info->block_group_cache_lock);
3351         return cache;
3352 }
3353
3354 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3355                             struct btrfs_trans_handle *trans,
3356                             struct btrfs_path *path)
3357 {
3358         struct btrfs_root *root = block_group->fs_info->tree_root;
3359         struct inode *inode = NULL;
3360         u64 alloc_hint = 0;
3361         int dcs = BTRFS_DC_ERROR;
3362         u64 num_pages = 0;
3363         int retries = 0;
3364         int ret = 0;
3365
3366         /*
3367          * If this block group is smaller than 100 megs don't bother caching the
3368          * block group.
3369          */
3370         if (block_group->key.offset < (100 * SZ_1M)) {
3371                 spin_lock(&block_group->lock);
3372                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3373                 spin_unlock(&block_group->lock);
3374                 return 0;
3375         }
3376
3377         if (trans->aborted)
3378                 return 0;
3379 again:
3380         inode = lookup_free_space_inode(root, block_group, path);
3381         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3382                 ret = PTR_ERR(inode);
3383                 btrfs_release_path(path);
3384                 goto out;
3385         }
3386
3387         if (IS_ERR(inode)) {
3388                 BUG_ON(retries);
3389                 retries++;
3390
3391                 if (block_group->ro)
3392                         goto out_free;
3393
3394                 ret = create_free_space_inode(root, trans, block_group, path);
3395                 if (ret)
3396                         goto out_free;
3397                 goto again;
3398         }
3399
3400         /* We've already setup this transaction, go ahead and exit */
3401         if (block_group->cache_generation == trans->transid &&
3402             i_size_read(inode)) {
3403                 dcs = BTRFS_DC_SETUP;
3404                 goto out_put;
3405         }
3406
3407         /*
3408          * We want to set the generation to 0, that way if anything goes wrong
3409          * from here on out we know not to trust this cache when we load up next
3410          * time.
3411          */
3412         BTRFS_I(inode)->generation = 0;
3413         ret = btrfs_update_inode(trans, root, inode);
3414         if (ret) {
3415                 /*
3416                  * So theoretically we could recover from this, simply set the
3417                  * super cache generation to 0 so we know to invalidate the
3418                  * cache, but then we'd have to keep track of the block groups
3419                  * that fail this way so we know we _have_ to reset this cache
3420                  * before the next commit or risk reading stale cache.  So to
3421                  * limit our exposure to horrible edge cases lets just abort the
3422                  * transaction, this only happens in really bad situations
3423                  * anyway.
3424                  */
3425                 btrfs_abort_transaction(trans, ret);
3426                 goto out_put;
3427         }
3428         WARN_ON(ret);
3429
3430         if (i_size_read(inode) > 0) {
3431                 ret = btrfs_check_trunc_cache_free_space(root,
3432                                         &root->fs_info->global_block_rsv);
3433                 if (ret)
3434                         goto out_put;
3435
3436                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3437                 if (ret)
3438                         goto out_put;
3439         }
3440
3441         spin_lock(&block_group->lock);
3442         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3443             !btrfs_test_opt(root->fs_info, SPACE_CACHE)) {
3444                 /*
3445                  * don't bother trying to write stuff out _if_
3446                  * a) we're not cached,
3447                  * b) we're with nospace_cache mount option.
3448                  */
3449                 dcs = BTRFS_DC_WRITTEN;
3450                 spin_unlock(&block_group->lock);
3451                 goto out_put;
3452         }
3453         spin_unlock(&block_group->lock);
3454
3455         /*
3456          * We hit an ENOSPC when setting up the cache in this transaction, just
3457          * skip doing the setup, we've already cleared the cache so we're safe.
3458          */
3459         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3460                 ret = -ENOSPC;
3461                 goto out_put;
3462         }
3463
3464         /*
3465          * Try to preallocate enough space based on how big the block group is.
3466          * Keep in mind this has to include any pinned space which could end up
3467          * taking up quite a bit since it's not folded into the other space
3468          * cache.
3469          */
3470         num_pages = div_u64(block_group->key.offset, SZ_256M);
3471         if (!num_pages)
3472                 num_pages = 1;
3473
3474         num_pages *= 16;
3475         num_pages *= PAGE_SIZE;
3476
3477         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3478         if (ret)
3479                 goto out_put;
3480
3481         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3482                                               num_pages, num_pages,
3483                                               &alloc_hint);
3484         /*
3485          * Our cache requires contiguous chunks so that we don't modify a bunch
3486          * of metadata or split extents when writing the cache out, which means
3487          * we can enospc if we are heavily fragmented in addition to just normal
3488          * out of space conditions.  So if we hit this just skip setting up any
3489          * other block groups for this transaction, maybe we'll unpin enough
3490          * space the next time around.
3491          */
3492         if (!ret)
3493                 dcs = BTRFS_DC_SETUP;
3494         else if (ret == -ENOSPC)
3495                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3496
3497 out_put:
3498         iput(inode);
3499 out_free:
3500         btrfs_release_path(path);
3501 out:
3502         spin_lock(&block_group->lock);
3503         if (!ret && dcs == BTRFS_DC_SETUP)
3504                 block_group->cache_generation = trans->transid;
3505         block_group->disk_cache_state = dcs;
3506         spin_unlock(&block_group->lock);
3507
3508         return ret;
3509 }
3510
3511 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3512                             struct btrfs_root *root)
3513 {
3514         struct btrfs_block_group_cache *cache, *tmp;
3515         struct btrfs_transaction *cur_trans = trans->transaction;
3516         struct btrfs_path *path;
3517
3518         if (list_empty(&cur_trans->dirty_bgs) ||
3519             !btrfs_test_opt(root->fs_info, SPACE_CACHE))
3520                 return 0;
3521
3522         path = btrfs_alloc_path();
3523         if (!path)
3524                 return -ENOMEM;
3525
3526         /* Could add new block groups, use _safe just in case */
3527         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3528                                  dirty_list) {
3529                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3530                         cache_save_setup(cache, trans, path);
3531         }
3532
3533         btrfs_free_path(path);
3534         return 0;
3535 }
3536
3537 /*
3538  * transaction commit does final block group cache writeback during a
3539  * critical section where nothing is allowed to change the FS.  This is
3540  * required in order for the cache to actually match the block group,
3541  * but can introduce a lot of latency into the commit.
3542  *
3543  * So, btrfs_start_dirty_block_groups is here to kick off block group
3544  * cache IO.  There's a chance we'll have to redo some of it if the
3545  * block group changes again during the commit, but it greatly reduces
3546  * the commit latency by getting rid of the easy block groups while
3547  * we're still allowing others to join the commit.
3548  */
3549 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3550                                    struct btrfs_root *root)
3551 {
3552         struct btrfs_block_group_cache *cache;
3553         struct btrfs_transaction *cur_trans = trans->transaction;
3554         int ret = 0;
3555         int should_put;
3556         struct btrfs_path *path = NULL;
3557         LIST_HEAD(dirty);
3558         struct list_head *io = &cur_trans->io_bgs;
3559         int num_started = 0;
3560         int loops = 0;
3561
3562         spin_lock(&cur_trans->dirty_bgs_lock);
3563         if (list_empty(&cur_trans->dirty_bgs)) {
3564                 spin_unlock(&cur_trans->dirty_bgs_lock);
3565                 return 0;
3566         }
3567         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3568         spin_unlock(&cur_trans->dirty_bgs_lock);
3569
3570 again:
3571         /*
3572          * make sure all the block groups on our dirty list actually
3573          * exist
3574          */
3575         btrfs_create_pending_block_groups(trans, root);
3576
3577         if (!path) {
3578                 path = btrfs_alloc_path();
3579                 if (!path)
3580                         return -ENOMEM;
3581         }
3582
3583         /*
3584          * cache_write_mutex is here only to save us from balance or automatic
3585          * removal of empty block groups deleting this block group while we are
3586          * writing out the cache
3587          */
3588         mutex_lock(&trans->transaction->cache_write_mutex);
3589         while (!list_empty(&dirty)) {
3590                 cache = list_first_entry(&dirty,
3591                                          struct btrfs_block_group_cache,
3592                                          dirty_list);
3593                 /*
3594                  * this can happen if something re-dirties a block
3595                  * group that is already under IO.  Just wait for it to
3596                  * finish and then do it all again
3597                  */
3598                 if (!list_empty(&cache->io_list)) {
3599                         list_del_init(&cache->io_list);
3600                         btrfs_wait_cache_io(root, trans, cache,
3601                                             &cache->io_ctl, path,
3602                                             cache->key.objectid);
3603                         btrfs_put_block_group(cache);
3604                 }
3605
3606
3607                 /*
3608                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3609                  * if it should update the cache_state.  Don't delete
3610                  * until after we wait.
3611                  *
3612                  * Since we're not running in the commit critical section
3613                  * we need the dirty_bgs_lock to protect from update_block_group
3614                  */
3615                 spin_lock(&cur_trans->dirty_bgs_lock);
3616                 list_del_init(&cache->dirty_list);
3617                 spin_unlock(&cur_trans->dirty_bgs_lock);
3618
3619                 should_put = 1;
3620
3621                 cache_save_setup(cache, trans, path);
3622
3623                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3624                         cache->io_ctl.inode = NULL;
3625                         ret = btrfs_write_out_cache(root, trans, cache, path);
3626                         if (ret == 0 && cache->io_ctl.inode) {
3627                                 num_started++;
3628                                 should_put = 0;
3629
3630                                 /*
3631                                  * the cache_write_mutex is protecting
3632                                  * the io_list
3633                                  */
3634                                 list_add_tail(&cache->io_list, io);
3635                         } else {
3636                                 /*
3637                                  * if we failed to write the cache, the
3638                                  * generation will be bad and life goes on
3639                                  */
3640                                 ret = 0;
3641                         }
3642                 }
3643                 if (!ret) {
3644                         ret = write_one_cache_group(trans, root, path, cache);
3645                         /*
3646                          * Our block group might still be attached to the list
3647                          * of new block groups in the transaction handle of some
3648                          * other task (struct btrfs_trans_handle->new_bgs). This
3649                          * means its block group item isn't yet in the extent
3650                          * tree. If this happens ignore the error, as we will
3651                          * try again later in the critical section of the
3652                          * transaction commit.
3653                          */
3654                         if (ret == -ENOENT) {
3655                                 ret = 0;
3656                                 spin_lock(&cur_trans->dirty_bgs_lock);
3657                                 if (list_empty(&cache->dirty_list)) {
3658                                         list_add_tail(&cache->dirty_list,
3659                                                       &cur_trans->dirty_bgs);
3660                                         btrfs_get_block_group(cache);
3661                                 }
3662                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3663                         } else if (ret) {
3664                                 btrfs_abort_transaction(trans, ret);
3665                         }
3666                 }
3667
3668                 /* if its not on the io list, we need to put the block group */
3669                 if (should_put)
3670                         btrfs_put_block_group(cache);
3671
3672                 if (ret)
3673                         break;
3674
3675                 /*
3676                  * Avoid blocking other tasks for too long. It might even save
3677                  * us from writing caches for block groups that are going to be
3678                  * removed.
3679                  */
3680                 mutex_unlock(&trans->transaction->cache_write_mutex);
3681                 mutex_lock(&trans->transaction->cache_write_mutex);
3682         }
3683         mutex_unlock(&trans->transaction->cache_write_mutex);
3684
3685         /*
3686          * go through delayed refs for all the stuff we've just kicked off
3687          * and then loop back (just once)
3688          */
3689         ret = btrfs_run_delayed_refs(trans, root, 0);
3690         if (!ret && loops == 0) {
3691                 loops++;
3692                 spin_lock(&cur_trans->dirty_bgs_lock);
3693                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3694                 /*
3695                  * dirty_bgs_lock protects us from concurrent block group
3696                  * deletes too (not just cache_write_mutex).
3697                  */
3698                 if (!list_empty(&dirty)) {
3699                         spin_unlock(&cur_trans->dirty_bgs_lock);
3700                         goto again;
3701                 }
3702                 spin_unlock(&cur_trans->dirty_bgs_lock);
3703         } else if (ret < 0) {
3704                 btrfs_cleanup_dirty_bgs(cur_trans, root);
3705         }
3706
3707         btrfs_free_path(path);
3708         return ret;
3709 }
3710
3711 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3712                                    struct btrfs_root *root)
3713 {
3714         struct btrfs_block_group_cache *cache;
3715         struct btrfs_transaction *cur_trans = trans->transaction;
3716         int ret = 0;
3717         int should_put;
3718         struct btrfs_path *path;
3719         struct list_head *io = &cur_trans->io_bgs;
3720         int num_started = 0;
3721
3722         path = btrfs_alloc_path();
3723         if (!path)
3724                 return -ENOMEM;
3725
3726         /*
3727          * Even though we are in the critical section of the transaction commit,
3728          * we can still have concurrent tasks adding elements to this
3729          * transaction's list of dirty block groups. These tasks correspond to
3730          * endio free space workers started when writeback finishes for a
3731          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3732          * allocate new block groups as a result of COWing nodes of the root
3733          * tree when updating the free space inode. The writeback for the space
3734          * caches is triggered by an earlier call to
3735          * btrfs_start_dirty_block_groups() and iterations of the following
3736          * loop.
3737          * Also we want to do the cache_save_setup first and then run the
3738          * delayed refs to make sure we have the best chance at doing this all
3739          * in one shot.
3740          */
3741         spin_lock(&cur_trans->dirty_bgs_lock);
3742         while (!list_empty(&cur_trans->dirty_bgs)) {
3743                 cache = list_first_entry(&cur_trans->dirty_bgs,
3744                                          struct btrfs_block_group_cache,
3745                                          dirty_list);
3746
3747                 /*
3748                  * this can happen if cache_save_setup re-dirties a block
3749                  * group that is already under IO.  Just wait for it to
3750                  * finish and then do it all again
3751                  */
3752                 if (!list_empty(&cache->io_list)) {
3753                         spin_unlock(&cur_trans->dirty_bgs_lock);
3754                         list_del_init(&cache->io_list);
3755                         btrfs_wait_cache_io(root, trans, cache,
3756                                             &cache->io_ctl, path,
3757                                             cache->key.objectid);
3758                         btrfs_put_block_group(cache);
3759                         spin_lock(&cur_trans->dirty_bgs_lock);
3760                 }
3761
3762                 /*
3763                  * don't remove from the dirty list until after we've waited
3764                  * on any pending IO
3765                  */
3766                 list_del_init(&cache->dirty_list);
3767                 spin_unlock(&cur_trans->dirty_bgs_lock);
3768                 should_put = 1;
3769
3770                 cache_save_setup(cache, trans, path);
3771
3772                 if (!ret)
3773                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3774
3775                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3776                         cache->io_ctl.inode = NULL;
3777                         ret = btrfs_write_out_cache(root, trans, cache, path);
3778                         if (ret == 0 && cache->io_ctl.inode) {
3779                                 num_started++;
3780                                 should_put = 0;
3781                                 list_add_tail(&cache->io_list, io);
3782                         } else {
3783                                 /*
3784                                  * if we failed to write the cache, the
3785                                  * generation will be bad and life goes on
3786                                  */
3787                                 ret = 0;
3788                         }
3789                 }
3790                 if (!ret) {
3791                         ret = write_one_cache_group(trans, root, path, cache);
3792                         /*
3793                          * One of the free space endio workers might have
3794                          * created a new block group while updating a free space
3795                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3796                          * and hasn't released its transaction handle yet, in
3797                          * which case the new block group is still attached to
3798                          * its transaction handle and its creation has not
3799                          * finished yet (no block group item in the extent tree
3800                          * yet, etc). If this is the case, wait for all free
3801                          * space endio workers to finish and retry. This is a
3802                          * a very rare case so no need for a more efficient and
3803                          * complex approach.
3804                          */
3805                         if (ret == -ENOENT) {
3806                                 wait_event(cur_trans->writer_wait,
3807                                    atomic_read(&cur_trans->num_writers) == 1);
3808                                 ret = write_one_cache_group(trans, root, path,
3809                                                             cache);
3810                         }
3811                         if (ret)
3812                                 btrfs_abort_transaction(trans, ret);
3813                 }
3814
3815                 /* if its not on the io list, we need to put the block group */
3816                 if (should_put)
3817                         btrfs_put_block_group(cache);
3818                 spin_lock(&cur_trans->dirty_bgs_lock);
3819         }
3820         spin_unlock(&cur_trans->dirty_bgs_lock);
3821
3822         while (!list_empty(io)) {
3823                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3824                                          io_list);
3825                 list_del_init(&cache->io_list);
3826                 btrfs_wait_cache_io(root, trans, cache,
3827                                     &cache->io_ctl, path, cache->key.objectid);
3828                 btrfs_put_block_group(cache);
3829         }
3830
3831         btrfs_free_path(path);
3832         return ret;
3833 }
3834
3835 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3836 {
3837         struct btrfs_block_group_cache *block_group;
3838         int readonly = 0;
3839
3840         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3841         if (!block_group || block_group->ro)
3842                 readonly = 1;
3843         if (block_group)
3844                 btrfs_put_block_group(block_group);
3845         return readonly;
3846 }
3847
3848 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3849 {
3850         struct btrfs_block_group_cache *bg;
3851         bool ret = true;
3852
3853         bg = btrfs_lookup_block_group(fs_info, bytenr);
3854         if (!bg)
3855                 return false;
3856
3857         spin_lock(&bg->lock);
3858         if (bg->ro)
3859                 ret = false;
3860         else
3861                 atomic_inc(&bg->nocow_writers);
3862         spin_unlock(&bg->lock);
3863
3864         /* no put on block group, done by btrfs_dec_nocow_writers */
3865         if (!ret)
3866                 btrfs_put_block_group(bg);
3867
3868         return ret;
3869
3870 }
3871
3872 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3873 {
3874         struct btrfs_block_group_cache *bg;
3875
3876         bg = btrfs_lookup_block_group(fs_info, bytenr);
3877         ASSERT(bg);
3878         if (atomic_dec_and_test(&bg->nocow_writers))
3879                 wake_up_atomic_t(&bg->nocow_writers);
3880         /*
3881          * Once for our lookup and once for the lookup done by a previous call
3882          * to btrfs_inc_nocow_writers()
3883          */
3884         btrfs_put_block_group(bg);
3885         btrfs_put_block_group(bg);
3886 }
3887
3888 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3889 {
3890         schedule();
3891         return 0;
3892 }
3893
3894 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3895 {
3896         wait_on_atomic_t(&bg->nocow_writers,
3897                          btrfs_wait_nocow_writers_atomic_t,
3898                          TASK_UNINTERRUPTIBLE);
3899 }
3900
3901 static const char *alloc_name(u64 flags)
3902 {
3903         switch (flags) {
3904         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3905                 return "mixed";
3906         case BTRFS_BLOCK_GROUP_METADATA:
3907                 return "metadata";
3908         case BTRFS_BLOCK_GROUP_DATA:
3909                 return "data";
3910         case BTRFS_BLOCK_GROUP_SYSTEM:
3911                 return "system";
3912         default:
3913                 WARN_ON(1);
3914                 return "invalid-combination";
3915         };
3916 }
3917
3918 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3919                              u64 total_bytes, u64 bytes_used,
3920                              u64 bytes_readonly,
3921                              struct btrfs_space_info **space_info)
3922 {
3923         struct btrfs_space_info *found;
3924         int i;
3925         int factor;
3926         int ret;
3927
3928         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3929                      BTRFS_BLOCK_GROUP_RAID10))
3930                 factor = 2;
3931         else
3932                 factor = 1;
3933
3934         found = __find_space_info(info, flags);
3935         if (found) {
3936                 spin_lock(&found->lock);
3937                 found->total_bytes += total_bytes;
3938                 found->disk_total += total_bytes * factor;
3939                 found->bytes_used += bytes_used;
3940                 found->disk_used += bytes_used * factor;
3941                 found->bytes_readonly += bytes_readonly;
3942                 if (total_bytes > 0)
3943                         found->full = 0;
3944                 space_info_add_new_bytes(info, found, total_bytes -
3945                                          bytes_used - bytes_readonly);
3946                 spin_unlock(&found->lock);
3947                 *space_info = found;
3948                 return 0;
3949         }
3950         found = kzalloc(sizeof(*found), GFP_NOFS);
3951         if (!found)
3952                 return -ENOMEM;
3953
3954         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3955         if (ret) {
3956                 kfree(found);
3957                 return ret;
3958         }
3959
3960         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3961                 INIT_LIST_HEAD(&found->block_groups[i]);
3962         init_rwsem(&found->groups_sem);
3963         spin_lock_init(&found->lock);
3964         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3965         found->total_bytes = total_bytes;
3966         found->disk_total = total_bytes * factor;
3967         found->bytes_used = bytes_used;
3968         found->disk_used = bytes_used * factor;
3969         found->bytes_pinned = 0;
3970         found->bytes_reserved = 0;
3971         found->bytes_readonly = bytes_readonly;
3972         found->bytes_may_use = 0;
3973         found->full = 0;
3974         found->max_extent_size = 0;
3975         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3976         found->chunk_alloc = 0;
3977         found->flush = 0;
3978         init_waitqueue_head(&found->wait);
3979         INIT_LIST_HEAD(&found->ro_bgs);
3980         INIT_LIST_HEAD(&found->tickets);
3981         INIT_LIST_HEAD(&found->priority_tickets);
3982
3983         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3984                                     info->space_info_kobj, "%s",
3985                                     alloc_name(found->flags));
3986         if (ret) {
3987                 kfree(found);
3988                 return ret;
3989         }
3990
3991         *space_info = found;
3992         list_add_rcu(&found->list, &info->space_info);
3993         if (flags & BTRFS_BLOCK_GROUP_DATA)
3994                 info->data_sinfo = found;
3995
3996         return ret;
3997 }
3998
3999 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4000 {
4001         u64 extra_flags = chunk_to_extended(flags) &
4002                                 BTRFS_EXTENDED_PROFILE_MASK;
4003
4004         write_seqlock(&fs_info->profiles_lock);
4005         if (flags & BTRFS_BLOCK_GROUP_DATA)
4006                 fs_info->avail_data_alloc_bits |= extra_flags;
4007         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4008                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4009         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4010                 fs_info->avail_system_alloc_bits |= extra_flags;
4011         write_sequnlock(&fs_info->profiles_lock);
4012 }
4013
4014 /*
4015  * returns target flags in extended format or 0 if restripe for this
4016  * chunk_type is not in progress
4017  *
4018  * should be called with either volume_mutex or balance_lock held
4019  */
4020 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4021 {
4022         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4023         u64 target = 0;
4024
4025         if (!bctl)
4026                 return 0;
4027
4028         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4029             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4030                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4031         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4032                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4033                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4034         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4035                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4036                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4037         }
4038
4039         return target;
4040 }
4041
4042 /*
4043  * @flags: available profiles in extended format (see ctree.h)
4044  *
4045  * Returns reduced profile in chunk format.  If profile changing is in
4046  * progress (either running or paused) picks the target profile (if it's
4047  * already available), otherwise falls back to plain reducing.
4048  */
4049 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4050 {
4051         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4052         u64 target;
4053         u64 raid_type;
4054         u64 allowed = 0;
4055
4056         /*
4057          * see if restripe for this chunk_type is in progress, if so
4058          * try to reduce to the target profile
4059          */
4060         spin_lock(&root->fs_info->balance_lock);
4061         target = get_restripe_target(root->fs_info, flags);
4062         if (target) {
4063                 /* pick target profile only if it's already available */
4064                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4065                         spin_unlock(&root->fs_info->balance_lock);
4066                         return extended_to_chunk(target);
4067                 }
4068         }
4069         spin_unlock(&root->fs_info->balance_lock);
4070
4071         /* First, mask out the RAID levels which aren't possible */
4072         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4073                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4074                         allowed |= btrfs_raid_group[raid_type];
4075         }
4076         allowed &= flags;
4077
4078         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4079                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4080         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4081                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4082         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4083                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4084         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4085                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4086         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4087                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4088
4089         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4090
4091         return extended_to_chunk(flags | allowed);
4092 }
4093
4094 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4095 {
4096         unsigned seq;
4097         u64 flags;
4098
4099         do {
4100                 flags = orig_flags;
4101                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4102
4103                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4104                         flags |= root->fs_info->avail_data_alloc_bits;
4105                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4106                         flags |= root->fs_info->avail_system_alloc_bits;
4107                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4108                         flags |= root->fs_info->avail_metadata_alloc_bits;
4109         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4110
4111         return btrfs_reduce_alloc_profile(root, flags);
4112 }
4113
4114 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4115 {
4116         u64 flags;
4117         u64 ret;
4118
4119         if (data)
4120                 flags = BTRFS_BLOCK_GROUP_DATA;
4121         else if (root == root->fs_info->chunk_root)
4122                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4123         else
4124                 flags = BTRFS_BLOCK_GROUP_METADATA;
4125
4126         ret = get_alloc_profile(root, flags);
4127         return ret;
4128 }
4129
4130 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4131 {
4132         struct btrfs_space_info *data_sinfo;
4133         struct btrfs_root *root = BTRFS_I(inode)->root;
4134         struct btrfs_fs_info *fs_info = root->fs_info;
4135         u64 used;
4136         int ret = 0;
4137         int need_commit = 2;
4138         int have_pinned_space;
4139
4140         /* make sure bytes are sectorsize aligned */
4141         bytes = ALIGN(bytes, root->sectorsize);
4142
4143         if (btrfs_is_free_space_inode(inode)) {
4144                 need_commit = 0;
4145                 ASSERT(current->journal_info);
4146         }
4147
4148         data_sinfo = fs_info->data_sinfo;
4149         if (!data_sinfo)
4150                 goto alloc;
4151
4152 again:
4153         /* make sure we have enough space to handle the data first */
4154         spin_lock(&data_sinfo->lock);
4155         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4156                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4157                 data_sinfo->bytes_may_use;
4158
4159         if (used + bytes > data_sinfo->total_bytes) {
4160                 struct btrfs_trans_handle *trans;
4161
4162                 /*
4163                  * if we don't have enough free bytes in this space then we need
4164                  * to alloc a new chunk.
4165                  */
4166                 if (!data_sinfo->full) {
4167                         u64 alloc_target;
4168
4169                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4170                         spin_unlock(&data_sinfo->lock);
4171 alloc:
4172                         alloc_target = btrfs_get_alloc_profile(root, 1);
4173                         /*
4174                          * It is ugly that we don't call nolock join
4175                          * transaction for the free space inode case here.
4176                          * But it is safe because we only do the data space
4177                          * reservation for the free space cache in the
4178                          * transaction context, the common join transaction
4179                          * just increase the counter of the current transaction
4180                          * handler, doesn't try to acquire the trans_lock of
4181                          * the fs.
4182                          */
4183                         trans = btrfs_join_transaction(root);
4184                         if (IS_ERR(trans))
4185                                 return PTR_ERR(trans);
4186
4187                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4188                                              alloc_target,
4189                                              CHUNK_ALLOC_NO_FORCE);
4190                         btrfs_end_transaction(trans, root);
4191                         if (ret < 0) {
4192                                 if (ret != -ENOSPC)
4193                                         return ret;
4194                                 else {
4195                                         have_pinned_space = 1;
4196                                         goto commit_trans;
4197                                 }
4198                         }
4199
4200                         if (!data_sinfo)
4201                                 data_sinfo = fs_info->data_sinfo;
4202
4203                         goto again;
4204                 }
4205
4206                 /*
4207                  * If we don't have enough pinned space to deal with this
4208                  * allocation, and no removed chunk in current transaction,
4209                  * don't bother committing the transaction.
4210                  */
4211                 have_pinned_space = percpu_counter_compare(
4212                         &data_sinfo->total_bytes_pinned,
4213                         used + bytes - data_sinfo->total_bytes);
4214                 spin_unlock(&data_sinfo->lock);
4215
4216                 /* commit the current transaction and try again */
4217 commit_trans:
4218                 if (need_commit &&
4219                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4220                         need_commit--;
4221
4222                         if (need_commit > 0) {
4223                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4224                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4225                         }
4226
4227                         trans = btrfs_join_transaction(root);
4228                         if (IS_ERR(trans))
4229                                 return PTR_ERR(trans);
4230                         if (have_pinned_space >= 0 ||
4231                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4232                                      &trans->transaction->flags) ||
4233                             need_commit > 0) {
4234                                 ret = btrfs_commit_transaction(trans, root);
4235                                 if (ret)
4236                                         return ret;
4237                                 /*
4238                                  * The cleaner kthread might still be doing iput
4239                                  * operations. Wait for it to finish so that
4240                                  * more space is released.
4241                                  */
4242                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4243                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4244                                 goto again;
4245                         } else {
4246                                 btrfs_end_transaction(trans, root);
4247                         }
4248                 }
4249
4250                 trace_btrfs_space_reservation(root->fs_info,
4251                                               "space_info:enospc",
4252                                               data_sinfo->flags, bytes, 1);
4253                 return -ENOSPC;
4254         }
4255         data_sinfo->bytes_may_use += bytes;
4256         trace_btrfs_space_reservation(root->fs_info, "space_info",
4257                                       data_sinfo->flags, bytes, 1);
4258         spin_unlock(&data_sinfo->lock);
4259
4260         return ret;
4261 }
4262
4263 /*
4264  * New check_data_free_space() with ability for precious data reservation
4265  * Will replace old btrfs_check_data_free_space(), but for patch split,
4266  * add a new function first and then replace it.
4267  */
4268 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4269 {
4270         struct btrfs_root *root = BTRFS_I(inode)->root;
4271         int ret;
4272
4273         /* align the range */
4274         len = round_up(start + len, root->sectorsize) -
4275               round_down(start, root->sectorsize);
4276         start = round_down(start, root->sectorsize);
4277
4278         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4279         if (ret < 0)
4280                 return ret;
4281
4282         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4283         ret = btrfs_qgroup_reserve_data(inode, start, len);
4284         if (ret)
4285                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4286         return ret;
4287 }
4288
4289 /*
4290  * Called if we need to clear a data reservation for this inode
4291  * Normally in a error case.
4292  *
4293  * This one will *NOT* use accurate qgroup reserved space API, just for case
4294  * which we can't sleep and is sure it won't affect qgroup reserved space.
4295  * Like clear_bit_hook().
4296  */
4297 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4298                                             u64 len)
4299 {
4300         struct btrfs_root *root = BTRFS_I(inode)->root;
4301         struct btrfs_space_info *data_sinfo;
4302
4303         /* Make sure the range is aligned to sectorsize */
4304         len = round_up(start + len, root->sectorsize) -
4305               round_down(start, root->sectorsize);
4306         start = round_down(start, root->sectorsize);
4307
4308         data_sinfo = root->fs_info->data_sinfo;
4309         spin_lock(&data_sinfo->lock);
4310         if (WARN_ON(data_sinfo->bytes_may_use < len))
4311                 data_sinfo->bytes_may_use = 0;
4312         else
4313                 data_sinfo->bytes_may_use -= len;
4314         trace_btrfs_space_reservation(root->fs_info, "space_info",
4315                                       data_sinfo->flags, len, 0);
4316         spin_unlock(&data_sinfo->lock);
4317 }
4318
4319 /*
4320  * Called if we need to clear a data reservation for this inode
4321  * Normally in a error case.
4322  *
4323  * This one will handle the per-inode data rsv map for accurate reserved
4324  * space framework.
4325  */
4326 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4327 {
4328         btrfs_free_reserved_data_space_noquota(inode, start, len);
4329         btrfs_qgroup_free_data(inode, start, len);
4330 }
4331
4332 static void force_metadata_allocation(struct btrfs_fs_info *info)
4333 {
4334         struct list_head *head = &info->space_info;
4335         struct btrfs_space_info *found;
4336
4337         rcu_read_lock();
4338         list_for_each_entry_rcu(found, head, list) {
4339                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4340                         found->force_alloc = CHUNK_ALLOC_FORCE;
4341         }
4342         rcu_read_unlock();
4343 }
4344
4345 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4346 {
4347         return (global->size << 1);
4348 }
4349
4350 static int should_alloc_chunk(struct btrfs_root *root,
4351                               struct btrfs_space_info *sinfo, int force)
4352 {
4353         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4354         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4355         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4356         u64 thresh;
4357
4358         if (force == CHUNK_ALLOC_FORCE)
4359                 return 1;
4360
4361         /*
4362          * We need to take into account the global rsv because for all intents
4363          * and purposes it's used space.  Don't worry about locking the
4364          * global_rsv, it doesn't change except when the transaction commits.
4365          */
4366         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4367                 num_allocated += calc_global_rsv_need_space(global_rsv);
4368
4369         /*
4370          * in limited mode, we want to have some free space up to
4371          * about 1% of the FS size.
4372          */
4373         if (force == CHUNK_ALLOC_LIMITED) {
4374                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4375                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4376
4377                 if (num_bytes - num_allocated < thresh)
4378                         return 1;
4379         }
4380
4381         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4382                 return 0;
4383         return 1;
4384 }
4385
4386 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4387 {
4388         u64 num_dev;
4389
4390         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4391                     BTRFS_BLOCK_GROUP_RAID0 |
4392                     BTRFS_BLOCK_GROUP_RAID5 |
4393                     BTRFS_BLOCK_GROUP_RAID6))
4394                 num_dev = root->fs_info->fs_devices->rw_devices;
4395         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4396                 num_dev = 2;
4397         else
4398                 num_dev = 1;    /* DUP or single */
4399
4400         return num_dev;
4401 }
4402
4403 /*
4404  * If @is_allocation is true, reserve space in the system space info necessary
4405  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4406  * removing a chunk.
4407  */
4408 void check_system_chunk(struct btrfs_trans_handle *trans,
4409                         struct btrfs_root *root,
4410                         u64 type)
4411 {
4412         struct btrfs_space_info *info;
4413         u64 left;
4414         u64 thresh;
4415         int ret = 0;
4416         u64 num_devs;
4417
4418         /*
4419          * Needed because we can end up allocating a system chunk and for an
4420          * atomic and race free space reservation in the chunk block reserve.
4421          */
4422         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4423
4424         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4425         spin_lock(&info->lock);
4426         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4427                 info->bytes_reserved - info->bytes_readonly -
4428                 info->bytes_may_use;
4429         spin_unlock(&info->lock);
4430
4431         num_devs = get_profile_num_devs(root, type);
4432
4433         /* num_devs device items to update and 1 chunk item to add or remove */
4434         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4435                 btrfs_calc_trans_metadata_size(root, 1);
4436
4437         if (left < thresh && btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
4438                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4439                         left, thresh, type);
4440                 dump_space_info(root->fs_info, info, 0, 0);
4441         }
4442
4443         if (left < thresh) {
4444                 u64 flags;
4445
4446                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4447                 /*
4448                  * Ignore failure to create system chunk. We might end up not
4449                  * needing it, as we might not need to COW all nodes/leafs from
4450                  * the paths we visit in the chunk tree (they were already COWed
4451                  * or created in the current transaction for example).
4452                  */
4453                 ret = btrfs_alloc_chunk(trans, root, flags);
4454         }
4455
4456         if (!ret) {
4457                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4458                                           &root->fs_info->chunk_block_rsv,
4459                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4460                 if (!ret)
4461                         trans->chunk_bytes_reserved += thresh;
4462         }
4463 }
4464
4465 /*
4466  * If force is CHUNK_ALLOC_FORCE:
4467  *    - return 1 if it successfully allocates a chunk,
4468  *    - return errors including -ENOSPC otherwise.
4469  * If force is NOT CHUNK_ALLOC_FORCE:
4470  *    - return 0 if it doesn't need to allocate a new chunk,
4471  *    - return 1 if it successfully allocates a chunk,
4472  *    - return errors including -ENOSPC otherwise.
4473  */
4474 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4475                           struct btrfs_root *extent_root, u64 flags, int force)
4476 {
4477         struct btrfs_space_info *space_info;
4478         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4479         int wait_for_alloc = 0;
4480         int ret = 0;
4481
4482         /* Don't re-enter if we're already allocating a chunk */
4483         if (trans->allocating_chunk)
4484                 return -ENOSPC;
4485
4486         space_info = __find_space_info(extent_root->fs_info, flags);
4487         if (!space_info) {
4488                 ret = update_space_info(extent_root->fs_info, flags,
4489                                         0, 0, 0, &space_info);
4490                 BUG_ON(ret); /* -ENOMEM */
4491         }
4492         BUG_ON(!space_info); /* Logic error */
4493
4494 again:
4495         spin_lock(&space_info->lock);
4496         if (force < space_info->force_alloc)
4497                 force = space_info->force_alloc;
4498         if (space_info->full) {
4499                 if (should_alloc_chunk(extent_root, space_info, force))
4500                         ret = -ENOSPC;
4501                 else
4502                         ret = 0;
4503                 spin_unlock(&space_info->lock);
4504                 return ret;
4505         }
4506
4507         if (!should_alloc_chunk(extent_root, space_info, force)) {
4508                 spin_unlock(&space_info->lock);
4509                 return 0;
4510         } else if (space_info->chunk_alloc) {
4511                 wait_for_alloc = 1;
4512         } else {
4513                 space_info->chunk_alloc = 1;
4514         }
4515
4516         spin_unlock(&space_info->lock);
4517
4518         mutex_lock(&fs_info->chunk_mutex);
4519
4520         /*
4521          * The chunk_mutex is held throughout the entirety of a chunk
4522          * allocation, so once we've acquired the chunk_mutex we know that the
4523          * other guy is done and we need to recheck and see if we should
4524          * allocate.
4525          */
4526         if (wait_for_alloc) {
4527                 mutex_unlock(&fs_info->chunk_mutex);
4528                 wait_for_alloc = 0;
4529                 goto again;
4530         }
4531
4532         trans->allocating_chunk = true;
4533
4534         /*
4535          * If we have mixed data/metadata chunks we want to make sure we keep
4536          * allocating mixed chunks instead of individual chunks.
4537          */
4538         if (btrfs_mixed_space_info(space_info))
4539                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4540
4541         /*
4542          * if we're doing a data chunk, go ahead and make sure that
4543          * we keep a reasonable number of metadata chunks allocated in the
4544          * FS as well.
4545          */
4546         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4547                 fs_info->data_chunk_allocations++;
4548                 if (!(fs_info->data_chunk_allocations %
4549                       fs_info->metadata_ratio))
4550                         force_metadata_allocation(fs_info);
4551         }
4552
4553         /*
4554          * Check if we have enough space in SYSTEM chunk because we may need
4555          * to update devices.
4556          */
4557         check_system_chunk(trans, extent_root, flags);
4558
4559         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4560         trans->allocating_chunk = false;
4561
4562         spin_lock(&space_info->lock);
4563         if (ret < 0 && ret != -ENOSPC)
4564                 goto out;
4565         if (ret)
4566                 space_info->full = 1;
4567         else
4568                 ret = 1;
4569
4570         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4571 out:
4572         space_info->chunk_alloc = 0;
4573         spin_unlock(&space_info->lock);
4574         mutex_unlock(&fs_info->chunk_mutex);
4575         /*
4576          * When we allocate a new chunk we reserve space in the chunk block
4577          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4578          * add new nodes/leafs to it if we end up needing to do it when
4579          * inserting the chunk item and updating device items as part of the
4580          * second phase of chunk allocation, performed by
4581          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4582          * large number of new block groups to create in our transaction
4583          * handle's new_bgs list to avoid exhausting the chunk block reserve
4584          * in extreme cases - like having a single transaction create many new
4585          * block groups when starting to write out the free space caches of all
4586          * the block groups that were made dirty during the lifetime of the
4587          * transaction.
4588          */
4589         if (trans->can_flush_pending_bgs &&
4590             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4591                 btrfs_create_pending_block_groups(trans, extent_root);
4592                 btrfs_trans_release_chunk_metadata(trans);
4593         }
4594         return ret;
4595 }
4596
4597 static int can_overcommit(struct btrfs_root *root,
4598                           struct btrfs_space_info *space_info, u64 bytes,
4599                           enum btrfs_reserve_flush_enum flush)
4600 {
4601         struct btrfs_block_rsv *global_rsv;
4602         u64 profile;
4603         u64 space_size;
4604         u64 avail;
4605         u64 used;
4606
4607         /* Don't overcommit when in mixed mode. */
4608         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4609                 return 0;
4610
4611         BUG_ON(root->fs_info == NULL);
4612         global_rsv = &root->fs_info->global_block_rsv;
4613         profile = btrfs_get_alloc_profile(root, 0);
4614         used = space_info->bytes_used + space_info->bytes_reserved +
4615                 space_info->bytes_pinned + space_info->bytes_readonly;
4616
4617         /*
4618          * We only want to allow over committing if we have lots of actual space
4619          * free, but if we don't have enough space to handle the global reserve
4620          * space then we could end up having a real enospc problem when trying
4621          * to allocate a chunk or some other such important allocation.
4622          */
4623         spin_lock(&global_rsv->lock);
4624         space_size = calc_global_rsv_need_space(global_rsv);
4625         spin_unlock(&global_rsv->lock);
4626         if (used + space_size >= space_info->total_bytes)
4627                 return 0;
4628
4629         used += space_info->bytes_may_use;
4630
4631         spin_lock(&root->fs_info->free_chunk_lock);
4632         avail = root->fs_info->free_chunk_space;
4633         spin_unlock(&root->fs_info->free_chunk_lock);
4634
4635         /*
4636          * If we have dup, raid1 or raid10 then only half of the free
4637          * space is actually useable.  For raid56, the space info used
4638          * doesn't include the parity drive, so we don't have to
4639          * change the math
4640          */
4641         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4642                        BTRFS_BLOCK_GROUP_RAID1 |
4643                        BTRFS_BLOCK_GROUP_RAID10))
4644                 avail >>= 1;
4645
4646         /*
4647          * If we aren't flushing all things, let us overcommit up to
4648          * 1/2th of the space. If we can flush, don't let us overcommit
4649          * too much, let it overcommit up to 1/8 of the space.
4650          */
4651         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4652                 avail >>= 3;
4653         else
4654                 avail >>= 1;
4655
4656         if (used + bytes < space_info->total_bytes + avail)
4657                 return 1;
4658         return 0;
4659 }
4660
4661 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4662                                          unsigned long nr_pages, int nr_items)
4663 {
4664         struct super_block *sb = root->fs_info->sb;
4665
4666         if (down_read_trylock(&sb->s_umount)) {
4667                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4668                 up_read(&sb->s_umount);
4669         } else {
4670                 /*
4671                  * We needn't worry the filesystem going from r/w to r/o though
4672                  * we don't acquire ->s_umount mutex, because the filesystem
4673                  * should guarantee the delalloc inodes list be empty after
4674                  * the filesystem is readonly(all dirty pages are written to
4675                  * the disk).
4676                  */
4677                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4678                 if (!current->journal_info)
4679                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4680                                                  0, (u64)-1);
4681         }
4682 }
4683
4684 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4685 {
4686         u64 bytes;
4687         int nr;
4688
4689         bytes = btrfs_calc_trans_metadata_size(root, 1);
4690         nr = (int)div64_u64(to_reclaim, bytes);
4691         if (!nr)
4692                 nr = 1;
4693         return nr;
4694 }
4695
4696 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4697
4698 /*
4699  * shrink metadata reservation for delalloc
4700  */
4701 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4702                             bool wait_ordered)
4703 {
4704         struct btrfs_block_rsv *block_rsv;
4705         struct btrfs_space_info *space_info;
4706         struct btrfs_trans_handle *trans;
4707         u64 delalloc_bytes;
4708         u64 max_reclaim;
4709         long time_left;
4710         unsigned long nr_pages;
4711         int loops;
4712         int items;
4713         enum btrfs_reserve_flush_enum flush;
4714
4715         /* Calc the number of the pages we need flush for space reservation */
4716         items = calc_reclaim_items_nr(root, to_reclaim);
4717         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4718
4719         trans = (struct btrfs_trans_handle *)current->journal_info;
4720         block_rsv = &root->fs_info->delalloc_block_rsv;
4721         space_info = block_rsv->space_info;
4722
4723         delalloc_bytes = percpu_counter_sum_positive(
4724                                                 &root->fs_info->delalloc_bytes);
4725         if (delalloc_bytes == 0) {
4726                 if (trans)
4727                         return;
4728                 if (wait_ordered)
4729                         btrfs_wait_ordered_roots(root->fs_info, items,
4730                                                  0, (u64)-1);
4731                 return;
4732         }
4733
4734         loops = 0;
4735         while (delalloc_bytes && loops < 3) {
4736                 max_reclaim = min(delalloc_bytes, to_reclaim);
4737                 nr_pages = max_reclaim >> PAGE_SHIFT;
4738                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4739                 /*
4740                  * We need to wait for the async pages to actually start before
4741                  * we do anything.
4742                  */
4743                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4744                 if (!max_reclaim)
4745                         goto skip_async;
4746
4747                 if (max_reclaim <= nr_pages)
4748                         max_reclaim = 0;
4749                 else
4750                         max_reclaim -= nr_pages;
4751
4752                 wait_event(root->fs_info->async_submit_wait,
4753                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4754                            (int)max_reclaim);
4755 skip_async:
4756                 if (!trans)
4757                         flush = BTRFS_RESERVE_FLUSH_ALL;
4758                 else
4759                         flush = BTRFS_RESERVE_NO_FLUSH;
4760                 spin_lock(&space_info->lock);
4761                 if (can_overcommit(root, space_info, orig, flush)) {
4762                         spin_unlock(&space_info->lock);
4763                         break;
4764                 }
4765                 if (list_empty(&space_info->tickets) &&
4766                     list_empty(&space_info->priority_tickets)) {
4767                         spin_unlock(&space_info->lock);
4768                         break;
4769                 }
4770                 spin_unlock(&space_info->lock);
4771
4772                 loops++;
4773                 if (wait_ordered && !trans) {
4774                         btrfs_wait_ordered_roots(root->fs_info, items,
4775                                                  0, (u64)-1);
4776                 } else {
4777                         time_left = schedule_timeout_killable(1);
4778                         if (time_left)
4779                                 break;
4780                 }
4781                 delalloc_bytes = percpu_counter_sum_positive(
4782                                                 &root->fs_info->delalloc_bytes);
4783         }
4784 }
4785
4786 /**
4787  * maybe_commit_transaction - possibly commit the transaction if its ok to
4788  * @root - the root we're allocating for
4789  * @bytes - the number of bytes we want to reserve
4790  * @force - force the commit
4791  *
4792  * This will check to make sure that committing the transaction will actually
4793  * get us somewhere and then commit the transaction if it does.  Otherwise it
4794  * will return -ENOSPC.
4795  */
4796 static int may_commit_transaction(struct btrfs_root *root,
4797                                   struct btrfs_space_info *space_info,
4798                                   u64 bytes, int force)
4799 {
4800         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4801         struct btrfs_trans_handle *trans;
4802
4803         trans = (struct btrfs_trans_handle *)current->journal_info;
4804         if (trans)
4805                 return -EAGAIN;
4806
4807         if (force)
4808                 goto commit;
4809
4810         /* See if there is enough pinned space to make this reservation */
4811         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4812                                    bytes) >= 0)
4813                 goto commit;
4814
4815         /*
4816          * See if there is some space in the delayed insertion reservation for
4817          * this reservation.
4818          */
4819         if (space_info != delayed_rsv->space_info)
4820                 return -ENOSPC;
4821
4822         spin_lock(&delayed_rsv->lock);
4823         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4824                                    bytes - delayed_rsv->size) >= 0) {
4825                 spin_unlock(&delayed_rsv->lock);
4826                 return -ENOSPC;
4827         }
4828         spin_unlock(&delayed_rsv->lock);
4829
4830 commit:
4831         trans = btrfs_join_transaction(root);
4832         if (IS_ERR(trans))
4833                 return -ENOSPC;
4834
4835         return btrfs_commit_transaction(trans, root);
4836 }
4837
4838 struct reserve_ticket {
4839         u64 bytes;
4840         int error;
4841         struct list_head list;
4842         wait_queue_head_t wait;
4843 };
4844
4845 static int flush_space(struct btrfs_root *root,
4846                        struct btrfs_space_info *space_info, u64 num_bytes,
4847                        u64 orig_bytes, int state)
4848 {
4849         struct btrfs_trans_handle *trans;
4850         int nr;
4851         int ret = 0;
4852
4853         switch (state) {
4854         case FLUSH_DELAYED_ITEMS_NR:
4855         case FLUSH_DELAYED_ITEMS:
4856                 if (state == FLUSH_DELAYED_ITEMS_NR)
4857                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4858                 else
4859                         nr = -1;
4860
4861                 trans = btrfs_join_transaction(root);
4862                 if (IS_ERR(trans)) {
4863                         ret = PTR_ERR(trans);
4864                         break;
4865                 }
4866                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4867                 btrfs_end_transaction(trans, root);
4868                 break;
4869         case FLUSH_DELALLOC:
4870         case FLUSH_DELALLOC_WAIT:
4871                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4872                                 state == FLUSH_DELALLOC_WAIT);
4873                 break;
4874         case ALLOC_CHUNK:
4875                 trans = btrfs_join_transaction(root);
4876                 if (IS_ERR(trans)) {
4877                         ret = PTR_ERR(trans);
4878                         break;
4879                 }
4880                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4881                                      btrfs_get_alloc_profile(root, 0),
4882                                      CHUNK_ALLOC_NO_FORCE);
4883                 btrfs_end_transaction(trans, root);
4884                 if (ret > 0 || ret == -ENOSPC)
4885                         ret = 0;
4886                 break;
4887         case COMMIT_TRANS:
4888                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4889                 break;
4890         default:
4891                 ret = -ENOSPC;
4892                 break;
4893         }
4894
4895         trace_btrfs_flush_space(root->fs_info, space_info->flags, num_bytes,
4896                                 orig_bytes, state, ret);
4897         return ret;
4898 }
4899
4900 static inline u64
4901 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4902                                  struct btrfs_space_info *space_info)
4903 {
4904         struct reserve_ticket *ticket;
4905         u64 used;
4906         u64 expected;
4907         u64 to_reclaim = 0;
4908
4909         list_for_each_entry(ticket, &space_info->tickets, list)
4910                 to_reclaim += ticket->bytes;
4911         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4912                 to_reclaim += ticket->bytes;
4913         if (to_reclaim)
4914                 return to_reclaim;
4915
4916         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4917         if (can_overcommit(root, space_info, to_reclaim,
4918                            BTRFS_RESERVE_FLUSH_ALL))
4919                 return 0;
4920
4921         used = space_info->bytes_used + space_info->bytes_reserved +
4922                space_info->bytes_pinned + space_info->bytes_readonly +
4923                space_info->bytes_may_use;
4924         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4925                 expected = div_factor_fine(space_info->total_bytes, 95);
4926         else
4927                 expected = div_factor_fine(space_info->total_bytes, 90);
4928
4929         if (used > expected)
4930                 to_reclaim = used - expected;
4931         else
4932                 to_reclaim = 0;
4933         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4934                                      space_info->bytes_reserved);
4935         return to_reclaim;
4936 }
4937
4938 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4939                                         struct btrfs_root *root, u64 used)
4940 {
4941         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4942
4943         /* If we're just plain full then async reclaim just slows us down. */
4944         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4945                 return 0;
4946
4947         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4948                 return 0;
4949
4950         return (used >= thresh && !btrfs_fs_closing(root->fs_info) &&
4951                 !test_bit(BTRFS_FS_STATE_REMOUNTING,
4952                           &root->fs_info->fs_state));
4953 }
4954
4955 static void wake_all_tickets(struct list_head *head)
4956 {
4957         struct reserve_ticket *ticket;
4958
4959         while (!list_empty(head)) {
4960                 ticket = list_first_entry(head, struct reserve_ticket, list);
4961                 list_del_init(&ticket->list);
4962                 ticket->error = -ENOSPC;
4963                 wake_up(&ticket->wait);
4964         }
4965 }
4966
4967 /*
4968  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4969  * will loop and continuously try to flush as long as we are making progress.
4970  * We count progress as clearing off tickets each time we have to loop.
4971  */
4972 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4973 {
4974         struct btrfs_fs_info *fs_info;
4975         struct btrfs_space_info *space_info;
4976         u64 to_reclaim;
4977         int flush_state;
4978         int commit_cycles = 0;
4979         u64 last_tickets_id;
4980
4981         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4982         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4983
4984         spin_lock(&space_info->lock);
4985         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4986                                                       space_info);
4987         if (!to_reclaim) {
4988                 space_info->flush = 0;
4989                 spin_unlock(&space_info->lock);
4990                 return;
4991         }
4992         last_tickets_id = space_info->tickets_id;
4993         spin_unlock(&space_info->lock);
4994
4995         flush_state = FLUSH_DELAYED_ITEMS_NR;
4996         do {
4997                 struct reserve_ticket *ticket;
4998                 int ret;
4999
5000                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
5001                             to_reclaim, flush_state);
5002                 spin_lock(&space_info->lock);
5003                 if (list_empty(&space_info->tickets)) {
5004                         space_info->flush = 0;
5005                         spin_unlock(&space_info->lock);
5006                         return;
5007                 }
5008                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5009                                                               space_info);
5010                 ticket = list_first_entry(&space_info->tickets,
5011                                           struct reserve_ticket, list);
5012                 if (last_tickets_id == space_info->tickets_id) {
5013                         flush_state++;
5014                 } else {
5015                         last_tickets_id = space_info->tickets_id;
5016                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5017                         if (commit_cycles)
5018                                 commit_cycles--;
5019                 }
5020
5021                 if (flush_state > COMMIT_TRANS) {
5022                         commit_cycles++;
5023                         if (commit_cycles > 2) {
5024                                 wake_all_tickets(&space_info->tickets);
5025                                 space_info->flush = 0;
5026                         } else {
5027                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5028                         }
5029                 }
5030                 spin_unlock(&space_info->lock);
5031         } while (flush_state <= COMMIT_TRANS);
5032 }
5033
5034 void btrfs_init_async_reclaim_work(struct work_struct *work)
5035 {
5036         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5037 }
5038
5039 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5040                                             struct btrfs_space_info *space_info,
5041                                             struct reserve_ticket *ticket)
5042 {
5043         u64 to_reclaim;
5044         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5045
5046         spin_lock(&space_info->lock);
5047         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5048                                                       space_info);
5049         if (!to_reclaim) {
5050                 spin_unlock(&space_info->lock);
5051                 return;
5052         }
5053         spin_unlock(&space_info->lock);
5054
5055         do {
5056                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5057                             to_reclaim, flush_state);
5058                 flush_state++;
5059                 spin_lock(&space_info->lock);
5060                 if (ticket->bytes == 0) {
5061                         spin_unlock(&space_info->lock);
5062                         return;
5063                 }
5064                 spin_unlock(&space_info->lock);
5065
5066                 /*
5067                  * Priority flushers can't wait on delalloc without
5068                  * deadlocking.
5069                  */
5070                 if (flush_state == FLUSH_DELALLOC ||
5071                     flush_state == FLUSH_DELALLOC_WAIT)
5072                         flush_state = ALLOC_CHUNK;
5073         } while (flush_state < COMMIT_TRANS);
5074 }
5075
5076 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5077                                struct btrfs_space_info *space_info,
5078                                struct reserve_ticket *ticket, u64 orig_bytes)
5079
5080 {
5081         DEFINE_WAIT(wait);
5082         int ret = 0;
5083
5084         spin_lock(&space_info->lock);
5085         while (ticket->bytes > 0 && ticket->error == 0) {
5086                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5087                 if (ret) {
5088                         ret = -EINTR;
5089                         break;
5090                 }
5091                 spin_unlock(&space_info->lock);
5092
5093                 schedule();
5094
5095                 finish_wait(&ticket->wait, &wait);
5096                 spin_lock(&space_info->lock);
5097         }
5098         if (!ret)
5099                 ret = ticket->error;
5100         if (!list_empty(&ticket->list))
5101                 list_del_init(&ticket->list);
5102         if (ticket->bytes && ticket->bytes < orig_bytes) {
5103                 u64 num_bytes = orig_bytes - ticket->bytes;
5104                 space_info->bytes_may_use -= num_bytes;
5105                 trace_btrfs_space_reservation(fs_info, "space_info",
5106                                               space_info->flags, num_bytes, 0);
5107         }
5108         spin_unlock(&space_info->lock);
5109
5110         return ret;
5111 }
5112
5113 /**
5114  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5115  * @root - the root we're allocating for
5116  * @space_info - the space info we want to allocate from
5117  * @orig_bytes - the number of bytes we want
5118  * @flush - whether or not we can flush to make our reservation
5119  *
5120  * This will reserve orig_bytes number of bytes from the space info associated
5121  * with the block_rsv.  If there is not enough space it will make an attempt to
5122  * flush out space to make room.  It will do this by flushing delalloc if
5123  * possible or committing the transaction.  If flush is 0 then no attempts to
5124  * regain reservations will be made and this will fail if there is not enough
5125  * space already.
5126  */
5127 static int __reserve_metadata_bytes(struct btrfs_root *root,
5128                                     struct btrfs_space_info *space_info,
5129                                     u64 orig_bytes,
5130                                     enum btrfs_reserve_flush_enum flush)
5131 {
5132         struct reserve_ticket ticket;
5133         u64 used;
5134         int ret = 0;
5135
5136         ASSERT(orig_bytes);
5137         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5138
5139         spin_lock(&space_info->lock);
5140         ret = -ENOSPC;
5141         used = space_info->bytes_used + space_info->bytes_reserved +
5142                 space_info->bytes_pinned + space_info->bytes_readonly +
5143                 space_info->bytes_may_use;
5144
5145         /*
5146          * If we have enough space then hooray, make our reservation and carry
5147          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5148          * If not things get more complicated.
5149          */
5150         if (used + orig_bytes <= space_info->total_bytes) {
5151                 space_info->bytes_may_use += orig_bytes;
5152                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5153                                               space_info->flags, orig_bytes,
5154                                               1);
5155                 ret = 0;
5156         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5157                 space_info->bytes_may_use += orig_bytes;
5158                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5159                                               space_info->flags, orig_bytes,
5160                                               1);
5161                 ret = 0;
5162         }
5163
5164         /*
5165          * If we couldn't make a reservation then setup our reservation ticket
5166          * and kick the async worker if it's not already running.
5167          *
5168          * If we are a priority flusher then we just need to add our ticket to
5169          * the list and we will do our own flushing further down.
5170          */
5171         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5172                 ticket.bytes = orig_bytes;
5173                 ticket.error = 0;
5174                 init_waitqueue_head(&ticket.wait);
5175                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5176                         list_add_tail(&ticket.list, &space_info->tickets);
5177                         if (!space_info->flush) {
5178                                 space_info->flush = 1;
5179                                 trace_btrfs_trigger_flush(root->fs_info,
5180                                                           space_info->flags,
5181                                                           orig_bytes, flush,
5182                                                           "enospc");
5183                                 queue_work(system_unbound_wq,
5184                                            &root->fs_info->async_reclaim_work);
5185                         }
5186                 } else {
5187                         list_add_tail(&ticket.list,
5188                                       &space_info->priority_tickets);
5189                 }
5190         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5191                 used += orig_bytes;
5192                 /*
5193                  * We will do the space reservation dance during log replay,
5194                  * which means we won't have fs_info->fs_root set, so don't do
5195                  * the async reclaim as we will panic.
5196                  */
5197                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags) &&
5198                     need_do_async_reclaim(space_info, root, used) &&
5199                     !work_busy(&root->fs_info->async_reclaim_work)) {
5200                         trace_btrfs_trigger_flush(root->fs_info,
5201                                                   space_info->flags,
5202                                                   orig_bytes, flush,
5203                                                   "preempt");
5204                         queue_work(system_unbound_wq,
5205                                    &root->fs_info->async_reclaim_work);
5206                 }
5207         }
5208         spin_unlock(&space_info->lock);
5209         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5210                 return ret;
5211
5212         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5213                 return wait_reserve_ticket(root->fs_info, space_info, &ticket,
5214                                            orig_bytes);
5215
5216         ret = 0;
5217         priority_reclaim_metadata_space(root->fs_info, space_info, &ticket);
5218         spin_lock(&space_info->lock);
5219         if (ticket.bytes) {
5220                 if (ticket.bytes < orig_bytes) {
5221                         u64 num_bytes = orig_bytes - ticket.bytes;
5222                         space_info->bytes_may_use -= num_bytes;
5223                         trace_btrfs_space_reservation(root->fs_info,
5224                                         "space_info", space_info->flags,
5225                                         num_bytes, 0);
5226
5227                 }
5228                 list_del_init(&ticket.list);
5229                 ret = -ENOSPC;
5230         }
5231         spin_unlock(&space_info->lock);
5232         ASSERT(list_empty(&ticket.list));
5233         return ret;
5234 }
5235
5236 /**
5237  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5238  * @root - the root we're allocating for
5239  * @block_rsv - the block_rsv we're allocating for
5240  * @orig_bytes - the number of bytes we want
5241  * @flush - whether or not we can flush to make our reservation
5242  *
5243  * This will reserve orgi_bytes number of bytes from the space info associated
5244  * with the block_rsv.  If there is not enough space it will make an attempt to
5245  * flush out space to make room.  It will do this by flushing delalloc if
5246  * possible or committing the transaction.  If flush is 0 then no attempts to
5247  * regain reservations will be made and this will fail if there is not enough
5248  * space already.
5249  */
5250 static int reserve_metadata_bytes(struct btrfs_root *root,
5251                                   struct btrfs_block_rsv *block_rsv,
5252                                   u64 orig_bytes,
5253                                   enum btrfs_reserve_flush_enum flush)
5254 {
5255         int ret;
5256
5257         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5258                                        flush);
5259         if (ret == -ENOSPC &&
5260             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5261                 struct btrfs_block_rsv *global_rsv =
5262                         &root->fs_info->global_block_rsv;
5263
5264                 if (block_rsv != global_rsv &&
5265                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5266                         ret = 0;
5267         }
5268         if (ret == -ENOSPC)
5269                 trace_btrfs_space_reservation(root->fs_info,
5270                                               "space_info:enospc",
5271                                               block_rsv->space_info->flags,
5272                                               orig_bytes, 1);
5273         return ret;
5274 }
5275
5276 static struct btrfs_block_rsv *get_block_rsv(
5277                                         const struct btrfs_trans_handle *trans,
5278                                         const struct btrfs_root *root)
5279 {
5280         struct btrfs_block_rsv *block_rsv = NULL;
5281
5282         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5283             (root == root->fs_info->csum_root && trans->adding_csums) ||
5284              (root == root->fs_info->uuid_root))
5285                 block_rsv = trans->block_rsv;
5286
5287         if (!block_rsv)
5288                 block_rsv = root->block_rsv;
5289
5290         if (!block_rsv)
5291                 block_rsv = &root->fs_info->empty_block_rsv;
5292
5293         return block_rsv;
5294 }
5295
5296 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5297                                u64 num_bytes)
5298 {
5299         int ret = -ENOSPC;
5300         spin_lock(&block_rsv->lock);
5301         if (block_rsv->reserved >= num_bytes) {
5302                 block_rsv->reserved -= num_bytes;
5303                 if (block_rsv->reserved < block_rsv->size)
5304                         block_rsv->full = 0;
5305                 ret = 0;
5306         }
5307         spin_unlock(&block_rsv->lock);
5308         return ret;
5309 }
5310
5311 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5312                                 u64 num_bytes, int update_size)
5313 {
5314         spin_lock(&block_rsv->lock);
5315         block_rsv->reserved += num_bytes;
5316         if (update_size)
5317                 block_rsv->size += num_bytes;
5318         else if (block_rsv->reserved >= block_rsv->size)
5319                 block_rsv->full = 1;
5320         spin_unlock(&block_rsv->lock);
5321 }
5322
5323 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5324                              struct btrfs_block_rsv *dest, u64 num_bytes,
5325                              int min_factor)
5326 {
5327         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5328         u64 min_bytes;
5329
5330         if (global_rsv->space_info != dest->space_info)
5331                 return -ENOSPC;
5332
5333         spin_lock(&global_rsv->lock);
5334         min_bytes = div_factor(global_rsv->size, min_factor);
5335         if (global_rsv->reserved < min_bytes + num_bytes) {
5336                 spin_unlock(&global_rsv->lock);
5337                 return -ENOSPC;
5338         }
5339         global_rsv->reserved -= num_bytes;
5340         if (global_rsv->reserved < global_rsv->size)
5341                 global_rsv->full = 0;
5342         spin_unlock(&global_rsv->lock);
5343
5344         block_rsv_add_bytes(dest, num_bytes, 1);
5345         return 0;
5346 }
5347
5348 /*
5349  * This is for space we already have accounted in space_info->bytes_may_use, so
5350  * basically when we're returning space from block_rsv's.
5351  */
5352 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5353                                      struct btrfs_space_info *space_info,
5354                                      u64 num_bytes)
5355 {
5356         struct reserve_ticket *ticket;
5357         struct list_head *head;
5358         u64 used;
5359         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5360         bool check_overcommit = false;
5361
5362         spin_lock(&space_info->lock);
5363         head = &space_info->priority_tickets;
5364
5365         /*
5366          * If we are over our limit then we need to check and see if we can
5367          * overcommit, and if we can't then we just need to free up our space
5368          * and not satisfy any requests.
5369          */
5370         used = space_info->bytes_used + space_info->bytes_reserved +
5371                 space_info->bytes_pinned + space_info->bytes_readonly +
5372                 space_info->bytes_may_use;
5373         if (used - num_bytes >= space_info->total_bytes)
5374                 check_overcommit = true;
5375 again:
5376         while (!list_empty(head) && num_bytes) {
5377                 ticket = list_first_entry(head, struct reserve_ticket,
5378                                           list);
5379                 /*
5380                  * We use 0 bytes because this space is already reserved, so
5381                  * adding the ticket space would be a double count.
5382                  */
5383                 if (check_overcommit &&
5384                     !can_overcommit(fs_info->extent_root, space_info, 0,
5385                                     flush))
5386                         break;
5387                 if (num_bytes >= ticket->bytes) {
5388                         list_del_init(&ticket->list);
5389                         num_bytes -= ticket->bytes;
5390                         ticket->bytes = 0;
5391                         space_info->tickets_id++;
5392                         wake_up(&ticket->wait);
5393                 } else {
5394                         ticket->bytes -= num_bytes;
5395                         num_bytes = 0;
5396                 }
5397         }
5398
5399         if (num_bytes && head == &space_info->priority_tickets) {
5400                 head = &space_info->tickets;
5401                 flush = BTRFS_RESERVE_FLUSH_ALL;
5402                 goto again;
5403         }
5404         space_info->bytes_may_use -= num_bytes;
5405         trace_btrfs_space_reservation(fs_info, "space_info",
5406                                       space_info->flags, num_bytes, 0);
5407         spin_unlock(&space_info->lock);
5408 }
5409
5410 /*
5411  * This is for newly allocated space that isn't accounted in
5412  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5413  * we use this helper.
5414  */
5415 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5416                                      struct btrfs_space_info *space_info,
5417                                      u64 num_bytes)
5418 {
5419         struct reserve_ticket *ticket;
5420         struct list_head *head = &space_info->priority_tickets;
5421
5422 again:
5423         while (!list_empty(head) && num_bytes) {
5424                 ticket = list_first_entry(head, struct reserve_ticket,
5425                                           list);
5426                 if (num_bytes >= ticket->bytes) {
5427                         trace_btrfs_space_reservation(fs_info, "space_info",
5428                                                       space_info->flags,
5429                                                       ticket->bytes, 1);
5430                         list_del_init(&ticket->list);
5431                         num_bytes -= ticket->bytes;
5432                         space_info->bytes_may_use += ticket->bytes;
5433                         ticket->bytes = 0;
5434                         space_info->tickets_id++;
5435                         wake_up(&ticket->wait);
5436                 } else {
5437                         trace_btrfs_space_reservation(fs_info, "space_info",
5438                                                       space_info->flags,
5439                                                       num_bytes, 1);
5440                         space_info->bytes_may_use += num_bytes;
5441                         ticket->bytes -= num_bytes;
5442                         num_bytes = 0;
5443                 }
5444         }
5445
5446         if (num_bytes && head == &space_info->priority_tickets) {
5447                 head = &space_info->tickets;
5448                 goto again;
5449         }
5450 }
5451
5452 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5453                                     struct btrfs_block_rsv *block_rsv,
5454                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5455 {
5456         struct btrfs_space_info *space_info = block_rsv->space_info;
5457
5458         spin_lock(&block_rsv->lock);
5459         if (num_bytes == (u64)-1)
5460                 num_bytes = block_rsv->size;
5461         block_rsv->size -= num_bytes;
5462         if (block_rsv->reserved >= block_rsv->size) {
5463                 num_bytes = block_rsv->reserved - block_rsv->size;
5464                 block_rsv->reserved = block_rsv->size;
5465                 block_rsv->full = 1;
5466         } else {
5467                 num_bytes = 0;
5468         }
5469         spin_unlock(&block_rsv->lock);
5470
5471         if (num_bytes > 0) {
5472                 if (dest) {
5473                         spin_lock(&dest->lock);
5474                         if (!dest->full) {
5475                                 u64 bytes_to_add;
5476
5477                                 bytes_to_add = dest->size - dest->reserved;
5478                                 bytes_to_add = min(num_bytes, bytes_to_add);
5479                                 dest->reserved += bytes_to_add;
5480                                 if (dest->reserved >= dest->size)
5481                                         dest->full = 1;
5482                                 num_bytes -= bytes_to_add;
5483                         }
5484                         spin_unlock(&dest->lock);
5485                 }
5486                 if (num_bytes)
5487                         space_info_add_old_bytes(fs_info, space_info,
5488                                                  num_bytes);
5489         }
5490 }
5491
5492 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5493                             struct btrfs_block_rsv *dst, u64 num_bytes,
5494                             int update_size)
5495 {
5496         int ret;
5497
5498         ret = block_rsv_use_bytes(src, num_bytes);
5499         if (ret)
5500                 return ret;
5501
5502         block_rsv_add_bytes(dst, num_bytes, update_size);
5503         return 0;
5504 }
5505
5506 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5507 {
5508         memset(rsv, 0, sizeof(*rsv));
5509         spin_lock_init(&rsv->lock);
5510         rsv->type = type;
5511 }
5512
5513 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5514                                               unsigned short type)
5515 {
5516         struct btrfs_block_rsv *block_rsv;
5517         struct btrfs_fs_info *fs_info = root->fs_info;
5518
5519         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5520         if (!block_rsv)
5521                 return NULL;
5522
5523         btrfs_init_block_rsv(block_rsv, type);
5524         block_rsv->space_info = __find_space_info(fs_info,
5525                                                   BTRFS_BLOCK_GROUP_METADATA);
5526         return block_rsv;
5527 }
5528
5529 void btrfs_free_block_rsv(struct btrfs_root *root,
5530                           struct btrfs_block_rsv *rsv)
5531 {
5532         if (!rsv)
5533                 return;
5534         btrfs_block_rsv_release(root, rsv, (u64)-1);
5535         kfree(rsv);
5536 }
5537
5538 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5539 {
5540         kfree(rsv);
5541 }
5542
5543 int btrfs_block_rsv_add(struct btrfs_root *root,
5544                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5545                         enum btrfs_reserve_flush_enum flush)
5546 {
5547         int ret;
5548
5549         if (num_bytes == 0)
5550                 return 0;
5551
5552         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5553         if (!ret) {
5554                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5555                 return 0;
5556         }
5557
5558         return ret;
5559 }
5560
5561 int btrfs_block_rsv_check(struct btrfs_root *root,
5562                           struct btrfs_block_rsv *block_rsv, int min_factor)
5563 {
5564         u64 num_bytes = 0;
5565         int ret = -ENOSPC;
5566
5567         if (!block_rsv)
5568                 return 0;
5569
5570         spin_lock(&block_rsv->lock);
5571         num_bytes = div_factor(block_rsv->size, min_factor);
5572         if (block_rsv->reserved >= num_bytes)
5573                 ret = 0;
5574         spin_unlock(&block_rsv->lock);
5575
5576         return ret;
5577 }
5578
5579 int btrfs_block_rsv_refill(struct btrfs_root *root,
5580                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5581                            enum btrfs_reserve_flush_enum flush)
5582 {
5583         u64 num_bytes = 0;
5584         int ret = -ENOSPC;
5585
5586         if (!block_rsv)
5587                 return 0;
5588
5589         spin_lock(&block_rsv->lock);
5590         num_bytes = min_reserved;
5591         if (block_rsv->reserved >= num_bytes)
5592                 ret = 0;
5593         else
5594                 num_bytes -= block_rsv->reserved;
5595         spin_unlock(&block_rsv->lock);
5596
5597         if (!ret)
5598                 return 0;
5599
5600         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5601         if (!ret) {
5602                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5603                 return 0;
5604         }
5605
5606         return ret;
5607 }
5608
5609 void btrfs_block_rsv_release(struct btrfs_root *root,
5610                              struct btrfs_block_rsv *block_rsv,
5611                              u64 num_bytes)
5612 {
5613         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5614         if (global_rsv == block_rsv ||
5615             block_rsv->space_info != global_rsv->space_info)
5616                 global_rsv = NULL;
5617         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5618                                 num_bytes);
5619 }
5620
5621 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5622 {
5623         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5624         struct btrfs_space_info *sinfo = block_rsv->space_info;
5625         u64 num_bytes;
5626
5627         /*
5628          * The global block rsv is based on the size of the extent tree, the
5629          * checksum tree and the root tree.  If the fs is empty we want to set
5630          * it to a minimal amount for safety.
5631          */
5632         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5633                 btrfs_root_used(&fs_info->csum_root->root_item) +
5634                 btrfs_root_used(&fs_info->tree_root->root_item);
5635         num_bytes = max_t(u64, num_bytes, SZ_16M);
5636
5637         spin_lock(&sinfo->lock);
5638         spin_lock(&block_rsv->lock);
5639
5640         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5641
5642         if (block_rsv->reserved < block_rsv->size) {
5643                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5644                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5645                         sinfo->bytes_may_use;
5646                 if (sinfo->total_bytes > num_bytes) {
5647                         num_bytes = sinfo->total_bytes - num_bytes;
5648                         num_bytes = min(num_bytes,
5649                                         block_rsv->size - block_rsv->reserved);
5650                         block_rsv->reserved += num_bytes;
5651                         sinfo->bytes_may_use += num_bytes;
5652                         trace_btrfs_space_reservation(fs_info, "space_info",
5653                                                       sinfo->flags, num_bytes,
5654                                                       1);
5655                 }
5656         } else if (block_rsv->reserved > block_rsv->size) {
5657                 num_bytes = block_rsv->reserved - block_rsv->size;
5658                 sinfo->bytes_may_use -= num_bytes;
5659                 trace_btrfs_space_reservation(fs_info, "space_info",
5660                                       sinfo->flags, num_bytes, 0);
5661                 block_rsv->reserved = block_rsv->size;
5662         }
5663
5664         if (block_rsv->reserved == block_rsv->size)
5665                 block_rsv->full = 1;
5666         else
5667                 block_rsv->full = 0;
5668
5669         spin_unlock(&block_rsv->lock);
5670         spin_unlock(&sinfo->lock);
5671 }
5672
5673 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5674 {
5675         struct btrfs_space_info *space_info;
5676
5677         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5678         fs_info->chunk_block_rsv.space_info = space_info;
5679
5680         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5681         fs_info->global_block_rsv.space_info = space_info;
5682         fs_info->delalloc_block_rsv.space_info = space_info;
5683         fs_info->trans_block_rsv.space_info = space_info;
5684         fs_info->empty_block_rsv.space_info = space_info;
5685         fs_info->delayed_block_rsv.space_info = space_info;
5686
5687         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5688         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5689         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5690         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5691         if (fs_info->quota_root)
5692                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5693         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5694
5695         update_global_block_rsv(fs_info);
5696 }
5697
5698 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5699 {
5700         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5701                                 (u64)-1);
5702         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5703         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5704         WARN_ON(fs_info->trans_block_rsv.size > 0);
5705         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5706         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5707         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5708         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5709         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5710 }
5711
5712 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5713                                   struct btrfs_root *root)
5714 {
5715         if (!trans->block_rsv)
5716                 return;
5717
5718         if (!trans->bytes_reserved)
5719                 return;
5720
5721         trace_btrfs_space_reservation(root->fs_info, "transaction",
5722                                       trans->transid, trans->bytes_reserved, 0);
5723         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5724         trans->bytes_reserved = 0;
5725 }
5726
5727 /*
5728  * To be called after all the new block groups attached to the transaction
5729  * handle have been created (btrfs_create_pending_block_groups()).
5730  */
5731 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5732 {
5733         struct btrfs_fs_info *fs_info = trans->fs_info;
5734
5735         if (!trans->chunk_bytes_reserved)
5736                 return;
5737
5738         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5739
5740         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5741                                 trans->chunk_bytes_reserved);
5742         trans->chunk_bytes_reserved = 0;
5743 }
5744
5745 /* Can only return 0 or -ENOSPC */
5746 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5747                                   struct inode *inode)
5748 {
5749         struct btrfs_root *root = BTRFS_I(inode)->root;
5750         /*
5751          * We always use trans->block_rsv here as we will have reserved space
5752          * for our orphan when starting the transaction, using get_block_rsv()
5753          * here will sometimes make us choose the wrong block rsv as we could be
5754          * doing a reloc inode for a non refcounted root.
5755          */
5756         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5757         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5758
5759         /*
5760          * We need to hold space in order to delete our orphan item once we've
5761          * added it, so this takes the reservation so we can release it later
5762          * when we are truly done with the orphan item.
5763          */
5764         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5765         trace_btrfs_space_reservation(root->fs_info, "orphan",
5766                                       btrfs_ino(inode), num_bytes, 1);
5767         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5768 }
5769
5770 void btrfs_orphan_release_metadata(struct inode *inode)
5771 {
5772         struct btrfs_root *root = BTRFS_I(inode)->root;
5773         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5774         trace_btrfs_space_reservation(root->fs_info, "orphan",
5775                                       btrfs_ino(inode), num_bytes, 0);
5776         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5777 }
5778
5779 /*
5780  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5781  * root: the root of the parent directory
5782  * rsv: block reservation
5783  * items: the number of items that we need do reservation
5784  * qgroup_reserved: used to return the reserved size in qgroup
5785  *
5786  * This function is used to reserve the space for snapshot/subvolume
5787  * creation and deletion. Those operations are different with the
5788  * common file/directory operations, they change two fs/file trees
5789  * and root tree, the number of items that the qgroup reserves is
5790  * different with the free space reservation. So we can not use
5791  * the space reservation mechanism in start_transaction().
5792  */
5793 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5794                                      struct btrfs_block_rsv *rsv,
5795                                      int items,
5796                                      u64 *qgroup_reserved,
5797                                      bool use_global_rsv)
5798 {
5799         u64 num_bytes;
5800         int ret;
5801         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5802
5803         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5804                 /* One for parent inode, two for dir entries */
5805                 num_bytes = 3 * root->nodesize;
5806                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5807                 if (ret)
5808                         return ret;
5809         } else {
5810                 num_bytes = 0;
5811         }
5812
5813         *qgroup_reserved = num_bytes;
5814
5815         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5816         rsv->space_info = __find_space_info(root->fs_info,
5817                                             BTRFS_BLOCK_GROUP_METADATA);
5818         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5819                                   BTRFS_RESERVE_FLUSH_ALL);
5820
5821         if (ret == -ENOSPC && use_global_rsv)
5822                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5823
5824         if (ret && *qgroup_reserved)
5825                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5826
5827         return ret;
5828 }
5829
5830 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5831                                       struct btrfs_block_rsv *rsv,
5832                                       u64 qgroup_reserved)
5833 {
5834         btrfs_block_rsv_release(root, rsv, (u64)-1);
5835 }
5836
5837 /**
5838  * drop_outstanding_extent - drop an outstanding extent
5839  * @inode: the inode we're dropping the extent for
5840  * @num_bytes: the number of bytes we're releasing.
5841  *
5842  * This is called when we are freeing up an outstanding extent, either called
5843  * after an error or after an extent is written.  This will return the number of
5844  * reserved extents that need to be freed.  This must be called with
5845  * BTRFS_I(inode)->lock held.
5846  */
5847 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5848 {
5849         unsigned drop_inode_space = 0;
5850         unsigned dropped_extents = 0;
5851         unsigned num_extents = 0;
5852
5853         num_extents = (unsigned)div64_u64(num_bytes +
5854                                           BTRFS_MAX_EXTENT_SIZE - 1,
5855                                           BTRFS_MAX_EXTENT_SIZE);
5856         ASSERT(num_extents);
5857         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5858         BTRFS_I(inode)->outstanding_extents -= num_extents;
5859
5860         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5861             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5862                                &BTRFS_I(inode)->runtime_flags))
5863                 drop_inode_space = 1;
5864
5865         /*
5866          * If we have more or the same amount of outstanding extents than we have
5867          * reserved then we need to leave the reserved extents count alone.
5868          */
5869         if (BTRFS_I(inode)->outstanding_extents >=
5870             BTRFS_I(inode)->reserved_extents)
5871                 return drop_inode_space;
5872
5873         dropped_extents = BTRFS_I(inode)->reserved_extents -
5874                 BTRFS_I(inode)->outstanding_extents;
5875         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5876         return dropped_extents + drop_inode_space;
5877 }
5878
5879 /**
5880  * calc_csum_metadata_size - return the amount of metadata space that must be
5881  *      reserved/freed for the given bytes.
5882  * @inode: the inode we're manipulating
5883  * @num_bytes: the number of bytes in question
5884  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5885  *
5886  * This adjusts the number of csum_bytes in the inode and then returns the
5887  * correct amount of metadata that must either be reserved or freed.  We
5888  * calculate how many checksums we can fit into one leaf and then divide the
5889  * number of bytes that will need to be checksumed by this value to figure out
5890  * how many checksums will be required.  If we are adding bytes then the number
5891  * may go up and we will return the number of additional bytes that must be
5892  * reserved.  If it is going down we will return the number of bytes that must
5893  * be freed.
5894  *
5895  * This must be called with BTRFS_I(inode)->lock held.
5896  */
5897 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5898                                    int reserve)
5899 {
5900         struct btrfs_root *root = BTRFS_I(inode)->root;
5901         u64 old_csums, num_csums;
5902
5903         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5904             BTRFS_I(inode)->csum_bytes == 0)
5905                 return 0;
5906
5907         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5908         if (reserve)
5909                 BTRFS_I(inode)->csum_bytes += num_bytes;
5910         else
5911                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5912         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5913
5914         /* No change, no need to reserve more */
5915         if (old_csums == num_csums)
5916                 return 0;
5917
5918         if (reserve)
5919                 return btrfs_calc_trans_metadata_size(root,
5920                                                       num_csums - old_csums);
5921
5922         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5923 }
5924
5925 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5926 {
5927         struct btrfs_root *root = BTRFS_I(inode)->root;
5928         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5929         u64 to_reserve = 0;
5930         u64 csum_bytes;
5931         unsigned nr_extents = 0;
5932         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5933         int ret = 0;
5934         bool delalloc_lock = true;
5935         u64 to_free = 0;
5936         unsigned dropped;
5937         bool release_extra = false;
5938
5939         /* If we are a free space inode we need to not flush since we will be in
5940          * the middle of a transaction commit.  We also don't need the delalloc
5941          * mutex since we won't race with anybody.  We need this mostly to make
5942          * lockdep shut its filthy mouth.
5943          *
5944          * If we have a transaction open (can happen if we call truncate_block
5945          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5946          */
5947         if (btrfs_is_free_space_inode(inode)) {
5948                 flush = BTRFS_RESERVE_NO_FLUSH;
5949                 delalloc_lock = false;
5950         } else if (current->journal_info) {
5951                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5952         }
5953
5954         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5955             btrfs_transaction_in_commit(root->fs_info))
5956                 schedule_timeout(1);
5957
5958         if (delalloc_lock)
5959                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5960
5961         num_bytes = ALIGN(num_bytes, root->sectorsize);
5962
5963         spin_lock(&BTRFS_I(inode)->lock);
5964         nr_extents = (unsigned)div64_u64(num_bytes +
5965                                          BTRFS_MAX_EXTENT_SIZE - 1,
5966                                          BTRFS_MAX_EXTENT_SIZE);
5967         BTRFS_I(inode)->outstanding_extents += nr_extents;
5968
5969         nr_extents = 0;
5970         if (BTRFS_I(inode)->outstanding_extents >
5971             BTRFS_I(inode)->reserved_extents)
5972                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5973                         BTRFS_I(inode)->reserved_extents;
5974
5975         /* We always want to reserve a slot for updating the inode. */
5976         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents + 1);
5977         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5978         csum_bytes = BTRFS_I(inode)->csum_bytes;
5979         spin_unlock(&BTRFS_I(inode)->lock);
5980
5981         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5982                 ret = btrfs_qgroup_reserve_meta(root,
5983                                 nr_extents * root->nodesize);
5984                 if (ret)
5985                         goto out_fail;
5986         }
5987
5988         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5989         if (unlikely(ret)) {
5990                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5991                 goto out_fail;
5992         }
5993
5994         spin_lock(&BTRFS_I(inode)->lock);
5995         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5996                              &BTRFS_I(inode)->runtime_flags)) {
5997                 to_reserve -= btrfs_calc_trans_metadata_size(root, 1);
5998                 release_extra = true;
5999         }
6000         BTRFS_I(inode)->reserved_extents += nr_extents;
6001         spin_unlock(&BTRFS_I(inode)->lock);
6002
6003         if (delalloc_lock)
6004                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6005
6006         if (to_reserve)
6007                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6008                                               btrfs_ino(inode), to_reserve, 1);
6009         if (release_extra)
6010                 btrfs_block_rsv_release(root, block_rsv,
6011                                         btrfs_calc_trans_metadata_size(root,
6012                                                                        1));
6013         return 0;
6014
6015 out_fail:
6016         spin_lock(&BTRFS_I(inode)->lock);
6017         dropped = drop_outstanding_extent(inode, num_bytes);
6018         /*
6019          * If the inodes csum_bytes is the same as the original
6020          * csum_bytes then we know we haven't raced with any free()ers
6021          * so we can just reduce our inodes csum bytes and carry on.
6022          */
6023         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6024                 calc_csum_metadata_size(inode, num_bytes, 0);
6025         } else {
6026                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6027                 u64 bytes;
6028
6029                 /*
6030                  * This is tricky, but first we need to figure out how much we
6031                  * freed from any free-ers that occurred during this
6032                  * reservation, so we reset ->csum_bytes to the csum_bytes
6033                  * before we dropped our lock, and then call the free for the
6034                  * number of bytes that were freed while we were trying our
6035                  * reservation.
6036                  */
6037                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6038                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6039                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6040
6041
6042                 /*
6043                  * Now we need to see how much we would have freed had we not
6044                  * been making this reservation and our ->csum_bytes were not
6045                  * artificially inflated.
6046                  */
6047                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6048                 bytes = csum_bytes - orig_csum_bytes;
6049                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6050
6051                 /*
6052                  * Now reset ->csum_bytes to what it should be.  If bytes is
6053                  * more than to_free then we would have freed more space had we
6054                  * not had an artificially high ->csum_bytes, so we need to free
6055                  * the remainder.  If bytes is the same or less then we don't
6056                  * need to do anything, the other free-ers did the correct
6057                  * thing.
6058                  */
6059                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6060                 if (bytes > to_free)
6061                         to_free = bytes - to_free;
6062                 else
6063                         to_free = 0;
6064         }
6065         spin_unlock(&BTRFS_I(inode)->lock);
6066         if (dropped)
6067                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6068
6069         if (to_free) {
6070                 btrfs_block_rsv_release(root, block_rsv, to_free);
6071                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6072                                               btrfs_ino(inode), to_free, 0);
6073         }
6074         if (delalloc_lock)
6075                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6076         return ret;
6077 }
6078
6079 /**
6080  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6081  * @inode: the inode to release the reservation for
6082  * @num_bytes: the number of bytes we're releasing
6083  *
6084  * This will release the metadata reservation for an inode.  This can be called
6085  * once we complete IO for a given set of bytes to release their metadata
6086  * reservations.
6087  */
6088 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6089 {
6090         struct btrfs_root *root = BTRFS_I(inode)->root;
6091         u64 to_free = 0;
6092         unsigned dropped;
6093
6094         num_bytes = ALIGN(num_bytes, root->sectorsize);
6095         spin_lock(&BTRFS_I(inode)->lock);
6096         dropped = drop_outstanding_extent(inode, num_bytes);
6097
6098         if (num_bytes)
6099                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6100         spin_unlock(&BTRFS_I(inode)->lock);
6101         if (dropped > 0)
6102                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6103
6104         if (btrfs_is_testing(root->fs_info))
6105                 return;
6106
6107         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6108                                       btrfs_ino(inode), to_free, 0);
6109
6110         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
6111                                 to_free);
6112 }
6113
6114 /**
6115  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6116  * delalloc
6117  * @inode: inode we're writing to
6118  * @start: start range we are writing to
6119  * @len: how long the range we are writing to
6120  *
6121  * This will do the following things
6122  *
6123  * o reserve space in data space info for num bytes
6124  *   and reserve precious corresponding qgroup space
6125  *   (Done in check_data_free_space)
6126  *
6127  * o reserve space for metadata space, based on the number of outstanding
6128  *   extents and how much csums will be needed
6129  *   also reserve metadata space in a per root over-reserve method.
6130  * o add to the inodes->delalloc_bytes
6131  * o add it to the fs_info's delalloc inodes list.
6132  *   (Above 3 all done in delalloc_reserve_metadata)
6133  *
6134  * Return 0 for success
6135  * Return <0 for error(-ENOSPC or -EQUOT)
6136  */
6137 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
6138 {
6139         int ret;
6140
6141         ret = btrfs_check_data_free_space(inode, start, len);
6142         if (ret < 0)
6143                 return ret;
6144         ret = btrfs_delalloc_reserve_metadata(inode, len);
6145         if (ret < 0)
6146                 btrfs_free_reserved_data_space(inode, start, len);
6147         return ret;
6148 }
6149
6150 /**
6151  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6152  * @inode: inode we're releasing space for
6153  * @start: start position of the space already reserved
6154  * @len: the len of the space already reserved
6155  *
6156  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6157  * called in the case that we don't need the metadata AND data reservations
6158  * anymore.  So if there is an error or we insert an inline extent.
6159  *
6160  * This function will release the metadata space that was not used and will
6161  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6162  * list if there are no delalloc bytes left.
6163  * Also it will handle the qgroup reserved space.
6164  */
6165 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
6166 {
6167         btrfs_delalloc_release_metadata(inode, len);
6168         btrfs_free_reserved_data_space(inode, start, len);
6169 }
6170
6171 static int update_block_group(struct btrfs_trans_handle *trans,
6172                               struct btrfs_root *root, u64 bytenr,
6173                               u64 num_bytes, int alloc)
6174 {
6175         struct btrfs_block_group_cache *cache = NULL;
6176         struct btrfs_fs_info *info = root->fs_info;
6177         u64 total = num_bytes;
6178         u64 old_val;
6179         u64 byte_in_group;
6180         int factor;
6181
6182         /* block accounting for super block */
6183         spin_lock(&info->delalloc_root_lock);
6184         old_val = btrfs_super_bytes_used(info->super_copy);
6185         if (alloc)
6186                 old_val += num_bytes;
6187         else
6188                 old_val -= num_bytes;
6189         btrfs_set_super_bytes_used(info->super_copy, old_val);
6190         spin_unlock(&info->delalloc_root_lock);
6191
6192         while (total) {
6193                 cache = btrfs_lookup_block_group(info, bytenr);
6194                 if (!cache)
6195                         return -ENOENT;
6196                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6197                                     BTRFS_BLOCK_GROUP_RAID1 |
6198                                     BTRFS_BLOCK_GROUP_RAID10))
6199                         factor = 2;
6200                 else
6201                         factor = 1;
6202                 /*
6203                  * If this block group has free space cache written out, we
6204                  * need to make sure to load it if we are removing space.  This
6205                  * is because we need the unpinning stage to actually add the
6206                  * space back to the block group, otherwise we will leak space.
6207                  */
6208                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6209                         cache_block_group(cache, 1);
6210
6211                 byte_in_group = bytenr - cache->key.objectid;
6212                 WARN_ON(byte_in_group > cache->key.offset);
6213
6214                 spin_lock(&cache->space_info->lock);
6215                 spin_lock(&cache->lock);
6216
6217                 if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
6218                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6219                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6220
6221                 old_val = btrfs_block_group_used(&cache->item);
6222                 num_bytes = min(total, cache->key.offset - byte_in_group);
6223                 if (alloc) {
6224                         old_val += num_bytes;
6225                         btrfs_set_block_group_used(&cache->item, old_val);
6226                         cache->reserved -= num_bytes;
6227                         cache->space_info->bytes_reserved -= num_bytes;
6228                         cache->space_info->bytes_used += num_bytes;
6229                         cache->space_info->disk_used += num_bytes * factor;
6230                         spin_unlock(&cache->lock);
6231                         spin_unlock(&cache->space_info->lock);
6232                 } else {
6233                         old_val -= num_bytes;
6234                         btrfs_set_block_group_used(&cache->item, old_val);
6235                         cache->pinned += num_bytes;
6236                         cache->space_info->bytes_pinned += num_bytes;
6237                         cache->space_info->bytes_used -= num_bytes;
6238                         cache->space_info->disk_used -= num_bytes * factor;
6239                         spin_unlock(&cache->lock);
6240                         spin_unlock(&cache->space_info->lock);
6241
6242                         trace_btrfs_space_reservation(root->fs_info, "pinned",
6243                                                       cache->space_info->flags,
6244                                                       num_bytes, 1);
6245                         set_extent_dirty(info->pinned_extents,
6246                                          bytenr, bytenr + num_bytes - 1,
6247                                          GFP_NOFS | __GFP_NOFAIL);
6248                 }
6249
6250                 spin_lock(&trans->transaction->dirty_bgs_lock);
6251                 if (list_empty(&cache->dirty_list)) {
6252                         list_add_tail(&cache->dirty_list,
6253                                       &trans->transaction->dirty_bgs);
6254                                 trans->transaction->num_dirty_bgs++;
6255                         btrfs_get_block_group(cache);
6256                 }
6257                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6258
6259                 /*
6260                  * No longer have used bytes in this block group, queue it for
6261                  * deletion. We do this after adding the block group to the
6262                  * dirty list to avoid races between cleaner kthread and space
6263                  * cache writeout.
6264                  */
6265                 if (!alloc && old_val == 0) {
6266                         spin_lock(&info->unused_bgs_lock);
6267                         if (list_empty(&cache->bg_list)) {
6268                                 btrfs_get_block_group(cache);
6269                                 list_add_tail(&cache->bg_list,
6270                                               &info->unused_bgs);
6271                         }
6272                         spin_unlock(&info->unused_bgs_lock);
6273                 }
6274
6275                 btrfs_put_block_group(cache);
6276                 total -= num_bytes;
6277                 bytenr += num_bytes;
6278         }
6279         return 0;
6280 }
6281
6282 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6283 {
6284         struct btrfs_block_group_cache *cache;
6285         u64 bytenr;
6286
6287         spin_lock(&root->fs_info->block_group_cache_lock);
6288         bytenr = root->fs_info->first_logical_byte;
6289         spin_unlock(&root->fs_info->block_group_cache_lock);
6290
6291         if (bytenr < (u64)-1)
6292                 return bytenr;
6293
6294         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6295         if (!cache)
6296                 return 0;
6297
6298         bytenr = cache->key.objectid;
6299         btrfs_put_block_group(cache);
6300
6301         return bytenr;
6302 }
6303
6304 static int pin_down_extent(struct btrfs_root *root,
6305                            struct btrfs_block_group_cache *cache,
6306                            u64 bytenr, u64 num_bytes, int reserved)
6307 {
6308         spin_lock(&cache->space_info->lock);
6309         spin_lock(&cache->lock);
6310         cache->pinned += num_bytes;
6311         cache->space_info->bytes_pinned += num_bytes;
6312         if (reserved) {
6313                 cache->reserved -= num_bytes;
6314                 cache->space_info->bytes_reserved -= num_bytes;
6315         }
6316         spin_unlock(&cache->lock);
6317         spin_unlock(&cache->space_info->lock);
6318
6319         trace_btrfs_space_reservation(root->fs_info, "pinned",
6320                                       cache->space_info->flags, num_bytes, 1);
6321         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6322                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6323         return 0;
6324 }
6325
6326 /*
6327  * this function must be called within transaction
6328  */
6329 int btrfs_pin_extent(struct btrfs_root *root,
6330                      u64 bytenr, u64 num_bytes, int reserved)
6331 {
6332         struct btrfs_block_group_cache *cache;
6333
6334         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6335         BUG_ON(!cache); /* Logic error */
6336
6337         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6338
6339         btrfs_put_block_group(cache);
6340         return 0;
6341 }
6342
6343 /*
6344  * this function must be called within transaction
6345  */
6346 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6347                                     u64 bytenr, u64 num_bytes)
6348 {
6349         struct btrfs_block_group_cache *cache;
6350         int ret;
6351
6352         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6353         if (!cache)
6354                 return -EINVAL;
6355
6356         /*
6357          * pull in the free space cache (if any) so that our pin
6358          * removes the free space from the cache.  We have load_only set
6359          * to one because the slow code to read in the free extents does check
6360          * the pinned extents.
6361          */
6362         cache_block_group(cache, 1);
6363
6364         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6365
6366         /* remove us from the free space cache (if we're there at all) */
6367         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6368         btrfs_put_block_group(cache);
6369         return ret;
6370 }
6371
6372 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6373 {
6374         int ret;
6375         struct btrfs_block_group_cache *block_group;
6376         struct btrfs_caching_control *caching_ctl;
6377
6378         block_group = btrfs_lookup_block_group(root->fs_info, start);
6379         if (!block_group)
6380                 return -EINVAL;
6381
6382         cache_block_group(block_group, 0);
6383         caching_ctl = get_caching_control(block_group);
6384
6385         if (!caching_ctl) {
6386                 /* Logic error */
6387                 BUG_ON(!block_group_cache_done(block_group));
6388                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6389         } else {
6390                 mutex_lock(&caching_ctl->mutex);
6391
6392                 if (start >= caching_ctl->progress) {
6393                         ret = add_excluded_extent(root, start, num_bytes);
6394                 } else if (start + num_bytes <= caching_ctl->progress) {
6395                         ret = btrfs_remove_free_space(block_group,
6396                                                       start, num_bytes);
6397                 } else {
6398                         num_bytes = caching_ctl->progress - start;
6399                         ret = btrfs_remove_free_space(block_group,
6400                                                       start, num_bytes);
6401                         if (ret)
6402                                 goto out_lock;
6403
6404                         num_bytes = (start + num_bytes) -
6405                                 caching_ctl->progress;
6406                         start = caching_ctl->progress;
6407                         ret = add_excluded_extent(root, start, num_bytes);
6408                 }
6409 out_lock:
6410                 mutex_unlock(&caching_ctl->mutex);
6411                 put_caching_control(caching_ctl);
6412         }
6413         btrfs_put_block_group(block_group);
6414         return ret;
6415 }
6416
6417 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6418                                  struct extent_buffer *eb)
6419 {
6420         struct btrfs_file_extent_item *item;
6421         struct btrfs_key key;
6422         int found_type;
6423         int i;
6424
6425         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6426                 return 0;
6427
6428         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6429                 btrfs_item_key_to_cpu(eb, &key, i);
6430                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6431                         continue;
6432                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6433                 found_type = btrfs_file_extent_type(eb, item);
6434                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6435                         continue;
6436                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6437                         continue;
6438                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6439                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6440                 __exclude_logged_extent(log, key.objectid, key.offset);
6441         }
6442
6443         return 0;
6444 }
6445
6446 static void
6447 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6448 {
6449         atomic_inc(&bg->reservations);
6450 }
6451
6452 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6453                                         const u64 start)
6454 {
6455         struct btrfs_block_group_cache *bg;
6456
6457         bg = btrfs_lookup_block_group(fs_info, start);
6458         ASSERT(bg);
6459         if (atomic_dec_and_test(&bg->reservations))
6460                 wake_up_atomic_t(&bg->reservations);
6461         btrfs_put_block_group(bg);
6462 }
6463
6464 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6465 {
6466         schedule();
6467         return 0;
6468 }
6469
6470 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6471 {
6472         struct btrfs_space_info *space_info = bg->space_info;
6473
6474         ASSERT(bg->ro);
6475
6476         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6477                 return;
6478
6479         /*
6480          * Our block group is read only but before we set it to read only,
6481          * some task might have had allocated an extent from it already, but it
6482          * has not yet created a respective ordered extent (and added it to a
6483          * root's list of ordered extents).
6484          * Therefore wait for any task currently allocating extents, since the
6485          * block group's reservations counter is incremented while a read lock
6486          * on the groups' semaphore is held and decremented after releasing
6487          * the read access on that semaphore and creating the ordered extent.
6488          */
6489         down_write(&space_info->groups_sem);
6490         up_write(&space_info->groups_sem);
6491
6492         wait_on_atomic_t(&bg->reservations,
6493                          btrfs_wait_bg_reservations_atomic_t,
6494                          TASK_UNINTERRUPTIBLE);
6495 }
6496
6497 /**
6498  * btrfs_add_reserved_bytes - update the block_group and space info counters
6499  * @cache:      The cache we are manipulating
6500  * @ram_bytes:  The number of bytes of file content, and will be same to
6501  *              @num_bytes except for the compress path.
6502  * @num_bytes:  The number of bytes in question
6503  * @delalloc:   The blocks are allocated for the delalloc write
6504  *
6505  * This is called by the allocator when it reserves space. Metadata
6506  * reservations should be called with RESERVE_ALLOC so we do the proper
6507  * ENOSPC accounting.  For data we handle the reservation through clearing the
6508  * delalloc bits in the io_tree.  We have to do this since we could end up
6509  * allocating less disk space for the amount of data we have reserved in the
6510  * case of compression.
6511  *
6512  * If this is a reservation and the block group has become read only we cannot
6513  * make the reservation and return -EAGAIN, otherwise this function always
6514  * succeeds.
6515  */
6516 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6517                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6518 {
6519         struct btrfs_space_info *space_info = cache->space_info;
6520         int ret = 0;
6521
6522         spin_lock(&space_info->lock);
6523         spin_lock(&cache->lock);
6524         if (cache->ro) {
6525                 ret = -EAGAIN;
6526         } else {
6527                 cache->reserved += num_bytes;
6528                 space_info->bytes_reserved += num_bytes;
6529
6530                 trace_btrfs_space_reservation(cache->fs_info,
6531                                 "space_info", space_info->flags,
6532                                 ram_bytes, 0);
6533                 space_info->bytes_may_use -= ram_bytes;
6534                 if (delalloc)
6535                         cache->delalloc_bytes += num_bytes;
6536         }
6537         spin_unlock(&cache->lock);
6538         spin_unlock(&space_info->lock);
6539         return ret;
6540 }
6541
6542 /**
6543  * btrfs_free_reserved_bytes - update the block_group and space info counters
6544  * @cache:      The cache we are manipulating
6545  * @num_bytes:  The number of bytes in question
6546  * @delalloc:   The blocks are allocated for the delalloc write
6547  *
6548  * This is called by somebody who is freeing space that was never actually used
6549  * on disk.  For example if you reserve some space for a new leaf in transaction
6550  * A and before transaction A commits you free that leaf, you call this with
6551  * reserve set to 0 in order to clear the reservation.
6552  */
6553
6554 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6555                                      u64 num_bytes, int delalloc)
6556 {
6557         struct btrfs_space_info *space_info = cache->space_info;
6558         int ret = 0;
6559
6560         spin_lock(&space_info->lock);
6561         spin_lock(&cache->lock);
6562         if (cache->ro)
6563                 space_info->bytes_readonly += num_bytes;
6564         cache->reserved -= num_bytes;
6565         space_info->bytes_reserved -= num_bytes;
6566
6567         if (delalloc)
6568                 cache->delalloc_bytes -= num_bytes;
6569         spin_unlock(&cache->lock);
6570         spin_unlock(&space_info->lock);
6571         return ret;
6572 }
6573 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6574                                 struct btrfs_root *root)
6575 {
6576         struct btrfs_fs_info *fs_info = root->fs_info;
6577         struct btrfs_caching_control *next;
6578         struct btrfs_caching_control *caching_ctl;
6579         struct btrfs_block_group_cache *cache;
6580
6581         down_write(&fs_info->commit_root_sem);
6582
6583         list_for_each_entry_safe(caching_ctl, next,
6584                                  &fs_info->caching_block_groups, list) {
6585                 cache = caching_ctl->block_group;
6586                 if (block_group_cache_done(cache)) {
6587                         cache->last_byte_to_unpin = (u64)-1;
6588                         list_del_init(&caching_ctl->list);
6589                         put_caching_control(caching_ctl);
6590                 } else {
6591                         cache->last_byte_to_unpin = caching_ctl->progress;
6592                 }
6593         }
6594
6595         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6596                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6597         else
6598                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6599
6600         up_write(&fs_info->commit_root_sem);
6601
6602         update_global_block_rsv(fs_info);
6603 }
6604
6605 /*
6606  * Returns the free cluster for the given space info and sets empty_cluster to
6607  * what it should be based on the mount options.
6608  */
6609 static struct btrfs_free_cluster *
6610 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6611                    u64 *empty_cluster)
6612 {
6613         struct btrfs_free_cluster *ret = NULL;
6614         bool ssd = btrfs_test_opt(root->fs_info, SSD);
6615
6616         *empty_cluster = 0;
6617         if (btrfs_mixed_space_info(space_info))
6618                 return ret;
6619
6620         if (ssd)
6621                 *empty_cluster = SZ_2M;
6622         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6623                 ret = &root->fs_info->meta_alloc_cluster;
6624                 if (!ssd)
6625                         *empty_cluster = SZ_64K;
6626         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6627                 ret = &root->fs_info->data_alloc_cluster;
6628         }
6629
6630         return ret;
6631 }
6632
6633 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6634                               const bool return_free_space)
6635 {
6636         struct btrfs_fs_info *fs_info = root->fs_info;
6637         struct btrfs_block_group_cache *cache = NULL;
6638         struct btrfs_space_info *space_info;
6639         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6640         struct btrfs_free_cluster *cluster = NULL;
6641         u64 len;
6642         u64 total_unpinned = 0;
6643         u64 empty_cluster = 0;
6644         bool readonly;
6645
6646         while (start <= end) {
6647                 readonly = false;
6648                 if (!cache ||
6649                     start >= cache->key.objectid + cache->key.offset) {
6650                         if (cache)
6651                                 btrfs_put_block_group(cache);
6652                         total_unpinned = 0;
6653                         cache = btrfs_lookup_block_group(fs_info, start);
6654                         BUG_ON(!cache); /* Logic error */
6655
6656                         cluster = fetch_cluster_info(root,
6657                                                      cache->space_info,
6658                                                      &empty_cluster);
6659                         empty_cluster <<= 1;
6660                 }
6661
6662                 len = cache->key.objectid + cache->key.offset - start;
6663                 len = min(len, end + 1 - start);
6664
6665                 if (start < cache->last_byte_to_unpin) {
6666                         len = min(len, cache->last_byte_to_unpin - start);
6667                         if (return_free_space)
6668                                 btrfs_add_free_space(cache, start, len);
6669                 }
6670
6671                 start += len;
6672                 total_unpinned += len;
6673                 space_info = cache->space_info;
6674
6675                 /*
6676                  * If this space cluster has been marked as fragmented and we've
6677                  * unpinned enough in this block group to potentially allow a
6678                  * cluster to be created inside of it go ahead and clear the
6679                  * fragmented check.
6680                  */
6681                 if (cluster && cluster->fragmented &&
6682                     total_unpinned > empty_cluster) {
6683                         spin_lock(&cluster->lock);
6684                         cluster->fragmented = 0;
6685                         spin_unlock(&cluster->lock);
6686                 }
6687
6688                 spin_lock(&space_info->lock);
6689                 spin_lock(&cache->lock);
6690                 cache->pinned -= len;
6691                 space_info->bytes_pinned -= len;
6692
6693                 trace_btrfs_space_reservation(fs_info, "pinned",
6694                                               space_info->flags, len, 0);
6695                 space_info->max_extent_size = 0;
6696                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6697                 if (cache->ro) {
6698                         space_info->bytes_readonly += len;
6699                         readonly = true;
6700                 }
6701                 spin_unlock(&cache->lock);
6702                 if (!readonly && return_free_space &&
6703                     global_rsv->space_info == space_info) {
6704                         u64 to_add = len;
6705                         WARN_ON(!return_free_space);
6706                         spin_lock(&global_rsv->lock);
6707                         if (!global_rsv->full) {
6708                                 to_add = min(len, global_rsv->size -
6709                                              global_rsv->reserved);
6710                                 global_rsv->reserved += to_add;
6711                                 space_info->bytes_may_use += to_add;
6712                                 if (global_rsv->reserved >= global_rsv->size)
6713                                         global_rsv->full = 1;
6714                                 trace_btrfs_space_reservation(fs_info,
6715                                                               "space_info",
6716                                                               space_info->flags,
6717                                                               to_add, 1);
6718                                 len -= to_add;
6719                         }
6720                         spin_unlock(&global_rsv->lock);
6721                         /* Add to any tickets we may have */
6722                         if (len)
6723                                 space_info_add_new_bytes(fs_info, space_info,
6724                                                          len);
6725                 }
6726                 spin_unlock(&space_info->lock);
6727         }
6728
6729         if (cache)
6730                 btrfs_put_block_group(cache);
6731         return 0;
6732 }
6733
6734 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6735                                struct btrfs_root *root)
6736 {
6737         struct btrfs_fs_info *fs_info = root->fs_info;
6738         struct btrfs_block_group_cache *block_group, *tmp;
6739         struct list_head *deleted_bgs;
6740         struct extent_io_tree *unpin;
6741         u64 start;
6742         u64 end;
6743         int ret;
6744
6745         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6746                 unpin = &fs_info->freed_extents[1];
6747         else
6748                 unpin = &fs_info->freed_extents[0];
6749
6750         while (!trans->aborted) {
6751                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6752                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6753                                             EXTENT_DIRTY, NULL);
6754                 if (ret) {
6755                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6756                         break;
6757                 }
6758
6759                 if (btrfs_test_opt(root->fs_info, DISCARD))
6760                         ret = btrfs_discard_extent(root, start,
6761                                                    end + 1 - start, NULL);
6762
6763                 clear_extent_dirty(unpin, start, end);
6764                 unpin_extent_range(root, start, end, true);
6765                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6766                 cond_resched();
6767         }
6768
6769         /*
6770          * Transaction is finished.  We don't need the lock anymore.  We
6771          * do need to clean up the block groups in case of a transaction
6772          * abort.
6773          */
6774         deleted_bgs = &trans->transaction->deleted_bgs;
6775         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6776                 u64 trimmed = 0;
6777
6778                 ret = -EROFS;
6779                 if (!trans->aborted)
6780                         ret = btrfs_discard_extent(root,
6781                                                    block_group->key.objectid,
6782                                                    block_group->key.offset,
6783                                                    &trimmed);
6784
6785                 list_del_init(&block_group->bg_list);
6786                 btrfs_put_block_group_trimming(block_group);
6787                 btrfs_put_block_group(block_group);
6788
6789                 if (ret) {
6790                         const char *errstr = btrfs_decode_error(ret);
6791                         btrfs_warn(fs_info,
6792                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6793                                    ret, errstr);
6794                 }
6795         }
6796
6797         return 0;
6798 }
6799
6800 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6801                              u64 owner, u64 root_objectid)
6802 {
6803         struct btrfs_space_info *space_info;
6804         u64 flags;
6805
6806         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6807                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6808                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6809                 else
6810                         flags = BTRFS_BLOCK_GROUP_METADATA;
6811         } else {
6812                 flags = BTRFS_BLOCK_GROUP_DATA;
6813         }
6814
6815         space_info = __find_space_info(fs_info, flags);
6816         BUG_ON(!space_info); /* Logic bug */
6817         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6818 }
6819
6820
6821 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6822                                 struct btrfs_root *root,
6823                                 struct btrfs_delayed_ref_node *node, u64 parent,
6824                                 u64 root_objectid, u64 owner_objectid,
6825                                 u64 owner_offset, int refs_to_drop,
6826                                 struct btrfs_delayed_extent_op *extent_op)
6827 {
6828         struct btrfs_key key;
6829         struct btrfs_path *path;
6830         struct btrfs_fs_info *info = root->fs_info;
6831         struct btrfs_root *extent_root = info->extent_root;
6832         struct extent_buffer *leaf;
6833         struct btrfs_extent_item *ei;
6834         struct btrfs_extent_inline_ref *iref;
6835         int ret;
6836         int is_data;
6837         int extent_slot = 0;
6838         int found_extent = 0;
6839         int num_to_del = 1;
6840         u32 item_size;
6841         u64 refs;
6842         u64 bytenr = node->bytenr;
6843         u64 num_bytes = node->num_bytes;
6844         int last_ref = 0;
6845         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6846                                                  SKINNY_METADATA);
6847
6848         path = btrfs_alloc_path();
6849         if (!path)
6850                 return -ENOMEM;
6851
6852         path->reada = READA_FORWARD;
6853         path->leave_spinning = 1;
6854
6855         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6856         BUG_ON(!is_data && refs_to_drop != 1);
6857
6858         if (is_data)
6859                 skinny_metadata = 0;
6860
6861         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6862                                     bytenr, num_bytes, parent,
6863                                     root_objectid, owner_objectid,
6864                                     owner_offset);
6865         if (ret == 0) {
6866                 extent_slot = path->slots[0];
6867                 while (extent_slot >= 0) {
6868                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6869                                               extent_slot);
6870                         if (key.objectid != bytenr)
6871                                 break;
6872                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6873                             key.offset == num_bytes) {
6874                                 found_extent = 1;
6875                                 break;
6876                         }
6877                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6878                             key.offset == owner_objectid) {
6879                                 found_extent = 1;
6880                                 break;
6881                         }
6882                         if (path->slots[0] - extent_slot > 5)
6883                                 break;
6884                         extent_slot--;
6885                 }
6886 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6887                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6888                 if (found_extent && item_size < sizeof(*ei))
6889                         found_extent = 0;
6890 #endif
6891                 if (!found_extent) {
6892                         BUG_ON(iref);
6893                         ret = remove_extent_backref(trans, extent_root, path,
6894                                                     NULL, refs_to_drop,
6895                                                     is_data, &last_ref);
6896                         if (ret) {
6897                                 btrfs_abort_transaction(trans, ret);
6898                                 goto out;
6899                         }
6900                         btrfs_release_path(path);
6901                         path->leave_spinning = 1;
6902
6903                         key.objectid = bytenr;
6904                         key.type = BTRFS_EXTENT_ITEM_KEY;
6905                         key.offset = num_bytes;
6906
6907                         if (!is_data && skinny_metadata) {
6908                                 key.type = BTRFS_METADATA_ITEM_KEY;
6909                                 key.offset = owner_objectid;
6910                         }
6911
6912                         ret = btrfs_search_slot(trans, extent_root,
6913                                                 &key, path, -1, 1);
6914                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6915                                 /*
6916                                  * Couldn't find our skinny metadata item,
6917                                  * see if we have ye olde extent item.
6918                                  */
6919                                 path->slots[0]--;
6920                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6921                                                       path->slots[0]);
6922                                 if (key.objectid == bytenr &&
6923                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6924                                     key.offset == num_bytes)
6925                                         ret = 0;
6926                         }
6927
6928                         if (ret > 0 && skinny_metadata) {
6929                                 skinny_metadata = false;
6930                                 key.objectid = bytenr;
6931                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6932                                 key.offset = num_bytes;
6933                                 btrfs_release_path(path);
6934                                 ret = btrfs_search_slot(trans, extent_root,
6935                                                         &key, path, -1, 1);
6936                         }
6937
6938                         if (ret) {
6939                                 btrfs_err(info,
6940                                           "umm, got %d back from search, was looking for %llu",
6941                                           ret, bytenr);
6942                                 if (ret > 0)
6943                                         btrfs_print_leaf(extent_root,
6944                                                          path->nodes[0]);
6945                         }
6946                         if (ret < 0) {
6947                                 btrfs_abort_transaction(trans, ret);
6948                                 goto out;
6949                         }
6950                         extent_slot = path->slots[0];
6951                 }
6952         } else if (WARN_ON(ret == -ENOENT)) {
6953                 btrfs_print_leaf(extent_root, path->nodes[0]);
6954                 btrfs_err(info,
6955                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6956                         bytenr, parent, root_objectid, owner_objectid,
6957                         owner_offset);
6958                 btrfs_abort_transaction(trans, ret);
6959                 goto out;
6960         } else {
6961                 btrfs_abort_transaction(trans, ret);
6962                 goto out;
6963         }
6964
6965         leaf = path->nodes[0];
6966         item_size = btrfs_item_size_nr(leaf, extent_slot);
6967 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6968         if (item_size < sizeof(*ei)) {
6969                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6970                 ret = convert_extent_item_v0(trans, extent_root, path,
6971                                              owner_objectid, 0);
6972                 if (ret < 0) {
6973                         btrfs_abort_transaction(trans, ret);
6974                         goto out;
6975                 }
6976
6977                 btrfs_release_path(path);
6978                 path->leave_spinning = 1;
6979
6980                 key.objectid = bytenr;
6981                 key.type = BTRFS_EXTENT_ITEM_KEY;
6982                 key.offset = num_bytes;
6983
6984                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6985                                         -1, 1);
6986                 if (ret) {
6987                         btrfs_err(info,
6988                                   "umm, got %d back from search, was looking for %llu",
6989                                 ret, bytenr);
6990                         btrfs_print_leaf(extent_root, path->nodes[0]);
6991                 }
6992                 if (ret < 0) {
6993                         btrfs_abort_transaction(trans, ret);
6994                         goto out;
6995                 }
6996
6997                 extent_slot = path->slots[0];
6998                 leaf = path->nodes[0];
6999                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7000         }
7001 #endif
7002         BUG_ON(item_size < sizeof(*ei));
7003         ei = btrfs_item_ptr(leaf, extent_slot,
7004                             struct btrfs_extent_item);
7005         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7006             key.type == BTRFS_EXTENT_ITEM_KEY) {
7007                 struct btrfs_tree_block_info *bi;
7008                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7009                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7010                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7011         }
7012
7013         refs = btrfs_extent_refs(leaf, ei);
7014         if (refs < refs_to_drop) {
7015                 btrfs_err(info,
7016                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7017                           refs_to_drop, refs, bytenr);
7018                 ret = -EINVAL;
7019                 btrfs_abort_transaction(trans, ret);
7020                 goto out;
7021         }
7022         refs -= refs_to_drop;
7023
7024         if (refs > 0) {
7025                 if (extent_op)
7026                         __run_delayed_extent_op(extent_op, leaf, ei);
7027                 /*
7028                  * In the case of inline back ref, reference count will
7029                  * be updated by remove_extent_backref
7030                  */
7031                 if (iref) {
7032                         BUG_ON(!found_extent);
7033                 } else {
7034                         btrfs_set_extent_refs(leaf, ei, refs);
7035                         btrfs_mark_buffer_dirty(leaf);
7036                 }
7037                 if (found_extent) {
7038                         ret = remove_extent_backref(trans, extent_root, path,
7039                                                     iref, refs_to_drop,
7040                                                     is_data, &last_ref);
7041                         if (ret) {
7042                                 btrfs_abort_transaction(trans, ret);
7043                                 goto out;
7044                         }
7045                 }
7046                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
7047                                  root_objectid);
7048         } else {
7049                 if (found_extent) {
7050                         BUG_ON(is_data && refs_to_drop !=
7051                                extent_data_ref_count(path, iref));
7052                         if (iref) {
7053                                 BUG_ON(path->slots[0] != extent_slot);
7054                         } else {
7055                                 BUG_ON(path->slots[0] != extent_slot + 1);
7056                                 path->slots[0] = extent_slot;
7057                                 num_to_del = 2;
7058                         }
7059                 }
7060
7061                 last_ref = 1;
7062                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7063                                       num_to_del);
7064                 if (ret) {
7065                         btrfs_abort_transaction(trans, ret);
7066                         goto out;
7067                 }
7068                 btrfs_release_path(path);
7069
7070                 if (is_data) {
7071                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
7072                         if (ret) {
7073                                 btrfs_abort_transaction(trans, ret);
7074                                 goto out;
7075                         }
7076                 }
7077
7078                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
7079                                              num_bytes);
7080                 if (ret) {
7081                         btrfs_abort_transaction(trans, ret);
7082                         goto out;
7083                 }
7084
7085                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
7086                 if (ret) {
7087                         btrfs_abort_transaction(trans, ret);
7088                         goto out;
7089                 }
7090         }
7091         btrfs_release_path(path);
7092
7093 out:
7094         btrfs_free_path(path);
7095         return ret;
7096 }
7097
7098 /*
7099  * when we free an block, it is possible (and likely) that we free the last
7100  * delayed ref for that extent as well.  This searches the delayed ref tree for
7101  * a given extent, and if there are no other delayed refs to be processed, it
7102  * removes it from the tree.
7103  */
7104 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7105                                       struct btrfs_root *root, u64 bytenr)
7106 {
7107         struct btrfs_delayed_ref_head *head;
7108         struct btrfs_delayed_ref_root *delayed_refs;
7109         int ret = 0;
7110
7111         delayed_refs = &trans->transaction->delayed_refs;
7112         spin_lock(&delayed_refs->lock);
7113         head = btrfs_find_delayed_ref_head(trans, bytenr);
7114         if (!head)
7115                 goto out_delayed_unlock;
7116
7117         spin_lock(&head->lock);
7118         if (!list_empty(&head->ref_list))
7119                 goto out;
7120
7121         if (head->extent_op) {
7122                 if (!head->must_insert_reserved)
7123                         goto out;
7124                 btrfs_free_delayed_extent_op(head->extent_op);
7125                 head->extent_op = NULL;
7126         }
7127
7128         /*
7129          * waiting for the lock here would deadlock.  If someone else has it
7130          * locked they are already in the process of dropping it anyway
7131          */
7132         if (!mutex_trylock(&head->mutex))
7133                 goto out;
7134
7135         /*
7136          * at this point we have a head with no other entries.  Go
7137          * ahead and process it.
7138          */
7139         head->node.in_tree = 0;
7140         rb_erase(&head->href_node, &delayed_refs->href_root);
7141
7142         atomic_dec(&delayed_refs->num_entries);
7143
7144         /*
7145          * we don't take a ref on the node because we're removing it from the
7146          * tree, so we just steal the ref the tree was holding.
7147          */
7148         delayed_refs->num_heads--;
7149         if (head->processing == 0)
7150                 delayed_refs->num_heads_ready--;
7151         head->processing = 0;
7152         spin_unlock(&head->lock);
7153         spin_unlock(&delayed_refs->lock);
7154
7155         BUG_ON(head->extent_op);
7156         if (head->must_insert_reserved)
7157                 ret = 1;
7158
7159         mutex_unlock(&head->mutex);
7160         btrfs_put_delayed_ref(&head->node);
7161         return ret;
7162 out:
7163         spin_unlock(&head->lock);
7164
7165 out_delayed_unlock:
7166         spin_unlock(&delayed_refs->lock);
7167         return 0;
7168 }
7169
7170 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7171                            struct btrfs_root *root,
7172                            struct extent_buffer *buf,
7173                            u64 parent, int last_ref)
7174 {
7175         int pin = 1;
7176         int ret;
7177
7178         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7179                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7180                                         buf->start, buf->len,
7181                                         parent, root->root_key.objectid,
7182                                         btrfs_header_level(buf),
7183                                         BTRFS_DROP_DELAYED_REF, NULL);
7184                 BUG_ON(ret); /* -ENOMEM */
7185         }
7186
7187         if (!last_ref)
7188                 return;
7189
7190         if (btrfs_header_generation(buf) == trans->transid) {
7191                 struct btrfs_block_group_cache *cache;
7192
7193                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7194                         ret = check_ref_cleanup(trans, root, buf->start);
7195                         if (!ret)
7196                                 goto out;
7197                 }
7198
7199                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
7200
7201                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7202                         pin_down_extent(root, cache, buf->start, buf->len, 1);
7203                         btrfs_put_block_group(cache);
7204                         goto out;
7205                 }
7206
7207                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7208
7209                 btrfs_add_free_space(cache, buf->start, buf->len);
7210                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7211                 btrfs_put_block_group(cache);
7212                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
7213                 pin = 0;
7214         }
7215 out:
7216         if (pin)
7217                 add_pinned_bytes(root->fs_info, buf->len,
7218                                  btrfs_header_level(buf),
7219                                  root->root_key.objectid);
7220
7221         /*
7222          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7223          * anymore.
7224          */
7225         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7226 }
7227
7228 /* Can return -ENOMEM */
7229 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7230                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7231                       u64 owner, u64 offset)
7232 {
7233         int ret;
7234         struct btrfs_fs_info *fs_info = root->fs_info;
7235
7236         if (btrfs_is_testing(fs_info))
7237                 return 0;
7238
7239         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
7240
7241         /*
7242          * tree log blocks never actually go into the extent allocation
7243          * tree, just update pinning info and exit early.
7244          */
7245         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7246                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7247                 /* unlocks the pinned mutex */
7248                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7249                 ret = 0;
7250         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7251                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7252                                         num_bytes,
7253                                         parent, root_objectid, (int)owner,
7254                                         BTRFS_DROP_DELAYED_REF, NULL);
7255         } else {
7256                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7257                                                 num_bytes,
7258                                                 parent, root_objectid, owner,
7259                                                 offset, 0,
7260                                                 BTRFS_DROP_DELAYED_REF, NULL);
7261         }
7262         return ret;
7263 }
7264
7265 /*
7266  * when we wait for progress in the block group caching, its because
7267  * our allocation attempt failed at least once.  So, we must sleep
7268  * and let some progress happen before we try again.
7269  *
7270  * This function will sleep at least once waiting for new free space to
7271  * show up, and then it will check the block group free space numbers
7272  * for our min num_bytes.  Another option is to have it go ahead
7273  * and look in the rbtree for a free extent of a given size, but this
7274  * is a good start.
7275  *
7276  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7277  * any of the information in this block group.
7278  */
7279 static noinline void
7280 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7281                                 u64 num_bytes)
7282 {
7283         struct btrfs_caching_control *caching_ctl;
7284
7285         caching_ctl = get_caching_control(cache);
7286         if (!caching_ctl)
7287                 return;
7288
7289         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7290                    (cache->free_space_ctl->free_space >= num_bytes));
7291
7292         put_caching_control(caching_ctl);
7293 }
7294
7295 static noinline int
7296 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7297 {
7298         struct btrfs_caching_control *caching_ctl;
7299         int ret = 0;
7300
7301         caching_ctl = get_caching_control(cache);
7302         if (!caching_ctl)
7303                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7304
7305         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7306         if (cache->cached == BTRFS_CACHE_ERROR)
7307                 ret = -EIO;
7308         put_caching_control(caching_ctl);
7309         return ret;
7310 }
7311
7312 int __get_raid_index(u64 flags)
7313 {
7314         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7315                 return BTRFS_RAID_RAID10;
7316         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7317                 return BTRFS_RAID_RAID1;
7318         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7319                 return BTRFS_RAID_DUP;
7320         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7321                 return BTRFS_RAID_RAID0;
7322         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7323                 return BTRFS_RAID_RAID5;
7324         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7325                 return BTRFS_RAID_RAID6;
7326
7327         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7328 }
7329
7330 int get_block_group_index(struct btrfs_block_group_cache *cache)
7331 {
7332         return __get_raid_index(cache->flags);
7333 }
7334
7335 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7336         [BTRFS_RAID_RAID10]     = "raid10",
7337         [BTRFS_RAID_RAID1]      = "raid1",
7338         [BTRFS_RAID_DUP]        = "dup",
7339         [BTRFS_RAID_RAID0]      = "raid0",
7340         [BTRFS_RAID_SINGLE]     = "single",
7341         [BTRFS_RAID_RAID5]      = "raid5",
7342         [BTRFS_RAID_RAID6]      = "raid6",
7343 };
7344
7345 static const char *get_raid_name(enum btrfs_raid_types type)
7346 {
7347         if (type >= BTRFS_NR_RAID_TYPES)
7348                 return NULL;
7349
7350         return btrfs_raid_type_names[type];
7351 }
7352
7353 enum btrfs_loop_type {
7354         LOOP_CACHING_NOWAIT = 0,
7355         LOOP_CACHING_WAIT = 1,
7356         LOOP_ALLOC_CHUNK = 2,
7357         LOOP_NO_EMPTY_SIZE = 3,
7358 };
7359
7360 static inline void
7361 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7362                        int delalloc)
7363 {
7364         if (delalloc)
7365                 down_read(&cache->data_rwsem);
7366 }
7367
7368 static inline void
7369 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7370                        int delalloc)
7371 {
7372         btrfs_get_block_group(cache);
7373         if (delalloc)
7374                 down_read(&cache->data_rwsem);
7375 }
7376
7377 static struct btrfs_block_group_cache *
7378 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7379                    struct btrfs_free_cluster *cluster,
7380                    int delalloc)
7381 {
7382         struct btrfs_block_group_cache *used_bg = NULL;
7383
7384         spin_lock(&cluster->refill_lock);
7385         while (1) {
7386                 used_bg = cluster->block_group;
7387                 if (!used_bg)
7388                         return NULL;
7389
7390                 if (used_bg == block_group)
7391                         return used_bg;
7392
7393                 btrfs_get_block_group(used_bg);
7394
7395                 if (!delalloc)
7396                         return used_bg;
7397
7398                 if (down_read_trylock(&used_bg->data_rwsem))
7399                         return used_bg;
7400
7401                 spin_unlock(&cluster->refill_lock);
7402
7403                 down_read(&used_bg->data_rwsem);
7404
7405                 spin_lock(&cluster->refill_lock);
7406                 if (used_bg == cluster->block_group)
7407                         return used_bg;
7408
7409                 up_read(&used_bg->data_rwsem);
7410                 btrfs_put_block_group(used_bg);
7411         }
7412 }
7413
7414 static inline void
7415 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7416                          int delalloc)
7417 {
7418         if (delalloc)
7419                 up_read(&cache->data_rwsem);
7420         btrfs_put_block_group(cache);
7421 }
7422
7423 /*
7424  * walks the btree of allocated extents and find a hole of a given size.
7425  * The key ins is changed to record the hole:
7426  * ins->objectid == start position
7427  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7428  * ins->offset == the size of the hole.
7429  * Any available blocks before search_start are skipped.
7430  *
7431  * If there is no suitable free space, we will record the max size of
7432  * the free space extent currently.
7433  */
7434 static noinline int find_free_extent(struct btrfs_root *orig_root,
7435                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7436                                 u64 hint_byte, struct btrfs_key *ins,
7437                                 u64 flags, int delalloc)
7438 {
7439         int ret = 0;
7440         struct btrfs_root *root = orig_root->fs_info->extent_root;
7441         struct btrfs_free_cluster *last_ptr = NULL;
7442         struct btrfs_block_group_cache *block_group = NULL;
7443         u64 search_start = 0;
7444         u64 max_extent_size = 0;
7445         u64 empty_cluster = 0;
7446         struct btrfs_space_info *space_info;
7447         int loop = 0;
7448         int index = __get_raid_index(flags);
7449         bool failed_cluster_refill = false;
7450         bool failed_alloc = false;
7451         bool use_cluster = true;
7452         bool have_caching_bg = false;
7453         bool orig_have_caching_bg = false;
7454         bool full_search = false;
7455
7456         WARN_ON(num_bytes < root->sectorsize);
7457         ins->type = BTRFS_EXTENT_ITEM_KEY;
7458         ins->objectid = 0;
7459         ins->offset = 0;
7460
7461         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7462
7463         space_info = __find_space_info(root->fs_info, flags);
7464         if (!space_info) {
7465                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7466                 return -ENOSPC;
7467         }
7468
7469         /*
7470          * If our free space is heavily fragmented we may not be able to make
7471          * big contiguous allocations, so instead of doing the expensive search
7472          * for free space, simply return ENOSPC with our max_extent_size so we
7473          * can go ahead and search for a more manageable chunk.
7474          *
7475          * If our max_extent_size is large enough for our allocation simply
7476          * disable clustering since we will likely not be able to find enough
7477          * space to create a cluster and induce latency trying.
7478          */
7479         if (unlikely(space_info->max_extent_size)) {
7480                 spin_lock(&space_info->lock);
7481                 if (space_info->max_extent_size &&
7482                     num_bytes > space_info->max_extent_size) {
7483                         ins->offset = space_info->max_extent_size;
7484                         spin_unlock(&space_info->lock);
7485                         return -ENOSPC;
7486                 } else if (space_info->max_extent_size) {
7487                         use_cluster = false;
7488                 }
7489                 spin_unlock(&space_info->lock);
7490         }
7491
7492         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7493         if (last_ptr) {
7494                 spin_lock(&last_ptr->lock);
7495                 if (last_ptr->block_group)
7496                         hint_byte = last_ptr->window_start;
7497                 if (last_ptr->fragmented) {
7498                         /*
7499                          * We still set window_start so we can keep track of the
7500                          * last place we found an allocation to try and save
7501                          * some time.
7502                          */
7503                         hint_byte = last_ptr->window_start;
7504                         use_cluster = false;
7505                 }
7506                 spin_unlock(&last_ptr->lock);
7507         }
7508
7509         search_start = max(search_start, first_logical_byte(root, 0));
7510         search_start = max(search_start, hint_byte);
7511         if (search_start == hint_byte) {
7512                 block_group = btrfs_lookup_block_group(root->fs_info,
7513                                                        search_start);
7514                 /*
7515                  * we don't want to use the block group if it doesn't match our
7516                  * allocation bits, or if its not cached.
7517                  *
7518                  * However if we are re-searching with an ideal block group
7519                  * picked out then we don't care that the block group is cached.
7520                  */
7521                 if (block_group && block_group_bits(block_group, flags) &&
7522                     block_group->cached != BTRFS_CACHE_NO) {
7523                         down_read(&space_info->groups_sem);
7524                         if (list_empty(&block_group->list) ||
7525                             block_group->ro) {
7526                                 /*
7527                                  * someone is removing this block group,
7528                                  * we can't jump into the have_block_group
7529                                  * target because our list pointers are not
7530                                  * valid
7531                                  */
7532                                 btrfs_put_block_group(block_group);
7533                                 up_read(&space_info->groups_sem);
7534                         } else {
7535                                 index = get_block_group_index(block_group);
7536                                 btrfs_lock_block_group(block_group, delalloc);
7537                                 goto have_block_group;
7538                         }
7539                 } else if (block_group) {
7540                         btrfs_put_block_group(block_group);
7541                 }
7542         }
7543 search:
7544         have_caching_bg = false;
7545         if (index == 0 || index == __get_raid_index(flags))
7546                 full_search = true;
7547         down_read(&space_info->groups_sem);
7548         list_for_each_entry(block_group, &space_info->block_groups[index],
7549                             list) {
7550                 u64 offset;
7551                 int cached;
7552
7553                 btrfs_grab_block_group(block_group, delalloc);
7554                 search_start = block_group->key.objectid;
7555
7556                 /*
7557                  * this can happen if we end up cycling through all the
7558                  * raid types, but we want to make sure we only allocate
7559                  * for the proper type.
7560                  */
7561                 if (!block_group_bits(block_group, flags)) {
7562                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7563                                 BTRFS_BLOCK_GROUP_RAID1 |
7564                                 BTRFS_BLOCK_GROUP_RAID5 |
7565                                 BTRFS_BLOCK_GROUP_RAID6 |
7566                                 BTRFS_BLOCK_GROUP_RAID10;
7567
7568                         /*
7569                          * if they asked for extra copies and this block group
7570                          * doesn't provide them, bail.  This does allow us to
7571                          * fill raid0 from raid1.
7572                          */
7573                         if ((flags & extra) && !(block_group->flags & extra))
7574                                 goto loop;
7575                 }
7576
7577 have_block_group:
7578                 cached = block_group_cache_done(block_group);
7579                 if (unlikely(!cached)) {
7580                         have_caching_bg = true;
7581                         ret = cache_block_group(block_group, 0);
7582                         BUG_ON(ret < 0);
7583                         ret = 0;
7584                 }
7585
7586                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7587                         goto loop;
7588                 if (unlikely(block_group->ro))
7589                         goto loop;
7590
7591                 /*
7592                  * Ok we want to try and use the cluster allocator, so
7593                  * lets look there
7594                  */
7595                 if (last_ptr && use_cluster) {
7596                         struct btrfs_block_group_cache *used_block_group;
7597                         unsigned long aligned_cluster;
7598                         /*
7599                          * the refill lock keeps out other
7600                          * people trying to start a new cluster
7601                          */
7602                         used_block_group = btrfs_lock_cluster(block_group,
7603                                                               last_ptr,
7604                                                               delalloc);
7605                         if (!used_block_group)
7606                                 goto refill_cluster;
7607
7608                         if (used_block_group != block_group &&
7609                             (used_block_group->ro ||
7610                              !block_group_bits(used_block_group, flags)))
7611                                 goto release_cluster;
7612
7613                         offset = btrfs_alloc_from_cluster(used_block_group,
7614                                                 last_ptr,
7615                                                 num_bytes,
7616                                                 used_block_group->key.objectid,
7617                                                 &max_extent_size);
7618                         if (offset) {
7619                                 /* we have a block, we're done */
7620                                 spin_unlock(&last_ptr->refill_lock);
7621                                 trace_btrfs_reserve_extent_cluster(root,
7622                                                 used_block_group,
7623                                                 search_start, num_bytes);
7624                                 if (used_block_group != block_group) {
7625                                         btrfs_release_block_group(block_group,
7626                                                                   delalloc);
7627                                         block_group = used_block_group;
7628                                 }
7629                                 goto checks;
7630                         }
7631
7632                         WARN_ON(last_ptr->block_group != used_block_group);
7633 release_cluster:
7634                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7635                          * set up a new clusters, so lets just skip it
7636                          * and let the allocator find whatever block
7637                          * it can find.  If we reach this point, we
7638                          * will have tried the cluster allocator
7639                          * plenty of times and not have found
7640                          * anything, so we are likely way too
7641                          * fragmented for the clustering stuff to find
7642                          * anything.
7643                          *
7644                          * However, if the cluster is taken from the
7645                          * current block group, release the cluster
7646                          * first, so that we stand a better chance of
7647                          * succeeding in the unclustered
7648                          * allocation.  */
7649                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7650                             used_block_group != block_group) {
7651                                 spin_unlock(&last_ptr->refill_lock);
7652                                 btrfs_release_block_group(used_block_group,
7653                                                           delalloc);
7654                                 goto unclustered_alloc;
7655                         }
7656
7657                         /*
7658                          * this cluster didn't work out, free it and
7659                          * start over
7660                          */
7661                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7662
7663                         if (used_block_group != block_group)
7664                                 btrfs_release_block_group(used_block_group,
7665                                                           delalloc);
7666 refill_cluster:
7667                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7668                                 spin_unlock(&last_ptr->refill_lock);
7669                                 goto unclustered_alloc;
7670                         }
7671
7672                         aligned_cluster = max_t(unsigned long,
7673                                                 empty_cluster + empty_size,
7674                                               block_group->full_stripe_len);
7675
7676                         /* allocate a cluster in this block group */
7677                         ret = btrfs_find_space_cluster(root, block_group,
7678                                                        last_ptr, search_start,
7679                                                        num_bytes,
7680                                                        aligned_cluster);
7681                         if (ret == 0) {
7682                                 /*
7683                                  * now pull our allocation out of this
7684                                  * cluster
7685                                  */
7686                                 offset = btrfs_alloc_from_cluster(block_group,
7687                                                         last_ptr,
7688                                                         num_bytes,
7689                                                         search_start,
7690                                                         &max_extent_size);
7691                                 if (offset) {
7692                                         /* we found one, proceed */
7693                                         spin_unlock(&last_ptr->refill_lock);
7694                                         trace_btrfs_reserve_extent_cluster(root,
7695                                                 block_group, search_start,
7696                                                 num_bytes);
7697                                         goto checks;
7698                                 }
7699                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7700                                    && !failed_cluster_refill) {
7701                                 spin_unlock(&last_ptr->refill_lock);
7702
7703                                 failed_cluster_refill = true;
7704                                 wait_block_group_cache_progress(block_group,
7705                                        num_bytes + empty_cluster + empty_size);
7706                                 goto have_block_group;
7707                         }
7708
7709                         /*
7710                          * at this point we either didn't find a cluster
7711                          * or we weren't able to allocate a block from our
7712                          * cluster.  Free the cluster we've been trying
7713                          * to use, and go to the next block group
7714                          */
7715                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7716                         spin_unlock(&last_ptr->refill_lock);
7717                         goto loop;
7718                 }
7719
7720 unclustered_alloc:
7721                 /*
7722                  * We are doing an unclustered alloc, set the fragmented flag so
7723                  * we don't bother trying to setup a cluster again until we get
7724                  * more space.
7725                  */
7726                 if (unlikely(last_ptr)) {
7727                         spin_lock(&last_ptr->lock);
7728                         last_ptr->fragmented = 1;
7729                         spin_unlock(&last_ptr->lock);
7730                 }
7731                 spin_lock(&block_group->free_space_ctl->tree_lock);
7732                 if (cached &&
7733                     block_group->free_space_ctl->free_space <
7734                     num_bytes + empty_cluster + empty_size) {
7735                         if (block_group->free_space_ctl->free_space >
7736                             max_extent_size)
7737                                 max_extent_size =
7738                                         block_group->free_space_ctl->free_space;
7739                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7740                         goto loop;
7741                 }
7742                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7743
7744                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7745                                                     num_bytes, empty_size,
7746                                                     &max_extent_size);
7747                 /*
7748                  * If we didn't find a chunk, and we haven't failed on this
7749                  * block group before, and this block group is in the middle of
7750                  * caching and we are ok with waiting, then go ahead and wait
7751                  * for progress to be made, and set failed_alloc to true.
7752                  *
7753                  * If failed_alloc is true then we've already waited on this
7754                  * block group once and should move on to the next block group.
7755                  */
7756                 if (!offset && !failed_alloc && !cached &&
7757                     loop > LOOP_CACHING_NOWAIT) {
7758                         wait_block_group_cache_progress(block_group,
7759                                                 num_bytes + empty_size);
7760                         failed_alloc = true;
7761                         goto have_block_group;
7762                 } else if (!offset) {
7763                         goto loop;
7764                 }
7765 checks:
7766                 search_start = ALIGN(offset, root->stripesize);
7767
7768                 /* move on to the next group */
7769                 if (search_start + num_bytes >
7770                     block_group->key.objectid + block_group->key.offset) {
7771                         btrfs_add_free_space(block_group, offset, num_bytes);
7772                         goto loop;
7773                 }
7774
7775                 if (offset < search_start)
7776                         btrfs_add_free_space(block_group, offset,
7777                                              search_start - offset);
7778                 BUG_ON(offset > search_start);
7779
7780                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7781                                 num_bytes, delalloc);
7782                 if (ret == -EAGAIN) {
7783                         btrfs_add_free_space(block_group, offset, num_bytes);
7784                         goto loop;
7785                 }
7786                 btrfs_inc_block_group_reservations(block_group);
7787
7788                 /* we are all good, lets return */
7789                 ins->objectid = search_start;
7790                 ins->offset = num_bytes;
7791
7792                 trace_btrfs_reserve_extent(orig_root, block_group,
7793                                            search_start, num_bytes);
7794                 btrfs_release_block_group(block_group, delalloc);
7795                 break;
7796 loop:
7797                 failed_cluster_refill = false;
7798                 failed_alloc = false;
7799                 BUG_ON(index != get_block_group_index(block_group));
7800                 btrfs_release_block_group(block_group, delalloc);
7801         }
7802         up_read(&space_info->groups_sem);
7803
7804         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7805                 && !orig_have_caching_bg)
7806                 orig_have_caching_bg = true;
7807
7808         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7809                 goto search;
7810
7811         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7812                 goto search;
7813
7814         /*
7815          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7816          *                      caching kthreads as we move along
7817          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7818          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7819          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7820          *                      again
7821          */
7822         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7823                 index = 0;
7824                 if (loop == LOOP_CACHING_NOWAIT) {
7825                         /*
7826                          * We want to skip the LOOP_CACHING_WAIT step if we
7827                          * don't have any uncached bgs and we've already done a
7828                          * full search through.
7829                          */
7830                         if (orig_have_caching_bg || !full_search)
7831                                 loop = LOOP_CACHING_WAIT;
7832                         else
7833                                 loop = LOOP_ALLOC_CHUNK;
7834                 } else {
7835                         loop++;
7836                 }
7837
7838                 if (loop == LOOP_ALLOC_CHUNK) {
7839                         struct btrfs_trans_handle *trans;
7840                         int exist = 0;
7841
7842                         trans = current->journal_info;
7843                         if (trans)
7844                                 exist = 1;
7845                         else
7846                                 trans = btrfs_join_transaction(root);
7847
7848                         if (IS_ERR(trans)) {
7849                                 ret = PTR_ERR(trans);
7850                                 goto out;
7851                         }
7852
7853                         ret = do_chunk_alloc(trans, root, flags,
7854                                              CHUNK_ALLOC_FORCE);
7855
7856                         /*
7857                          * If we can't allocate a new chunk we've already looped
7858                          * through at least once, move on to the NO_EMPTY_SIZE
7859                          * case.
7860                          */
7861                         if (ret == -ENOSPC)
7862                                 loop = LOOP_NO_EMPTY_SIZE;
7863
7864                         /*
7865                          * Do not bail out on ENOSPC since we
7866                          * can do more things.
7867                          */
7868                         if (ret < 0 && ret != -ENOSPC)
7869                                 btrfs_abort_transaction(trans, ret);
7870                         else
7871                                 ret = 0;
7872                         if (!exist)
7873                                 btrfs_end_transaction(trans, root);
7874                         if (ret)
7875                                 goto out;
7876                 }
7877
7878                 if (loop == LOOP_NO_EMPTY_SIZE) {
7879                         /*
7880                          * Don't loop again if we already have no empty_size and
7881                          * no empty_cluster.
7882                          */
7883                         if (empty_size == 0 &&
7884                             empty_cluster == 0) {
7885                                 ret = -ENOSPC;
7886                                 goto out;
7887                         }
7888                         empty_size = 0;
7889                         empty_cluster = 0;
7890                 }
7891
7892                 goto search;
7893         } else if (!ins->objectid) {
7894                 ret = -ENOSPC;
7895         } else if (ins->objectid) {
7896                 if (!use_cluster && last_ptr) {
7897                         spin_lock(&last_ptr->lock);
7898                         last_ptr->window_start = ins->objectid;
7899                         spin_unlock(&last_ptr->lock);
7900                 }
7901                 ret = 0;
7902         }
7903 out:
7904         if (ret == -ENOSPC) {
7905                 spin_lock(&space_info->lock);
7906                 space_info->max_extent_size = max_extent_size;
7907                 spin_unlock(&space_info->lock);
7908                 ins->offset = max_extent_size;
7909         }
7910         return ret;
7911 }
7912
7913 static void dump_space_info(struct btrfs_fs_info *fs_info,
7914                             struct btrfs_space_info *info, u64 bytes,
7915                             int dump_block_groups)
7916 {
7917         struct btrfs_block_group_cache *cache;
7918         int index = 0;
7919
7920         spin_lock(&info->lock);
7921         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7922                    info->flags,
7923                    info->total_bytes - info->bytes_used - info->bytes_pinned -
7924                    info->bytes_reserved - info->bytes_readonly -
7925                    info->bytes_may_use, (info->full) ? "" : "not ");
7926         btrfs_info(fs_info,
7927                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7928                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7929                 info->bytes_reserved, info->bytes_may_use,
7930                 info->bytes_readonly);
7931         spin_unlock(&info->lock);
7932
7933         if (!dump_block_groups)
7934                 return;
7935
7936         down_read(&info->groups_sem);
7937 again:
7938         list_for_each_entry(cache, &info->block_groups[index], list) {
7939                 spin_lock(&cache->lock);
7940                 btrfs_info(fs_info,
7941                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7942                         cache->key.objectid, cache->key.offset,
7943                         btrfs_block_group_used(&cache->item), cache->pinned,
7944                         cache->reserved, cache->ro ? "[readonly]" : "");
7945                 btrfs_dump_free_space(cache, bytes);
7946                 spin_unlock(&cache->lock);
7947         }
7948         if (++index < BTRFS_NR_RAID_TYPES)
7949                 goto again;
7950         up_read(&info->groups_sem);
7951 }
7952
7953 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7954                          u64 num_bytes, u64 min_alloc_size,
7955                          u64 empty_size, u64 hint_byte,
7956                          struct btrfs_key *ins, int is_data, int delalloc)
7957 {
7958         struct btrfs_fs_info *fs_info = root->fs_info;
7959         bool final_tried = num_bytes == min_alloc_size;
7960         u64 flags;
7961         int ret;
7962
7963         flags = btrfs_get_alloc_profile(root, is_data);
7964 again:
7965         WARN_ON(num_bytes < root->sectorsize);
7966         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
7967                                hint_byte, ins, flags, delalloc);
7968         if (!ret && !is_data) {
7969                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7970         } else if (ret == -ENOSPC) {
7971                 if (!final_tried && ins->offset) {
7972                         num_bytes = min(num_bytes >> 1, ins->offset);
7973                         num_bytes = round_down(num_bytes, root->sectorsize);
7974                         num_bytes = max(num_bytes, min_alloc_size);
7975                         ram_bytes = num_bytes;
7976                         if (num_bytes == min_alloc_size)
7977                                 final_tried = true;
7978                         goto again;
7979                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7980                         struct btrfs_space_info *sinfo;
7981
7982                         sinfo = __find_space_info(fs_info, flags);
7983                         btrfs_err(root->fs_info,
7984                                   "allocation failed flags %llu, wanted %llu",
7985                                   flags, num_bytes);
7986                         if (sinfo)
7987                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7988                 }
7989         }
7990
7991         return ret;
7992 }
7993
7994 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7995                                         u64 start, u64 len,
7996                                         int pin, int delalloc)
7997 {
7998         struct btrfs_block_group_cache *cache;
7999         int ret = 0;
8000
8001         cache = btrfs_lookup_block_group(root->fs_info, start);
8002         if (!cache) {
8003                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
8004                         start);
8005                 return -ENOSPC;
8006         }
8007
8008         if (pin)
8009                 pin_down_extent(root, cache, start, len, 1);
8010         else {
8011                 if (btrfs_test_opt(root->fs_info, DISCARD))
8012                         ret = btrfs_discard_extent(root, start, len, NULL);
8013                 btrfs_add_free_space(cache, start, len);
8014                 btrfs_free_reserved_bytes(cache, len, delalloc);
8015                 trace_btrfs_reserved_extent_free(root, start, len);
8016         }
8017
8018         btrfs_put_block_group(cache);
8019         return ret;
8020 }
8021
8022 int btrfs_free_reserved_extent(struct btrfs_root *root,
8023                                u64 start, u64 len, int delalloc)
8024 {
8025         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
8026 }
8027
8028 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
8029                                        u64 start, u64 len)
8030 {
8031         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
8032 }
8033
8034 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8035                                       struct btrfs_root *root,
8036                                       u64 parent, u64 root_objectid,
8037                                       u64 flags, u64 owner, u64 offset,
8038                                       struct btrfs_key *ins, int ref_mod)
8039 {
8040         int ret;
8041         struct btrfs_fs_info *fs_info = root->fs_info;
8042         struct btrfs_extent_item *extent_item;
8043         struct btrfs_extent_inline_ref *iref;
8044         struct btrfs_path *path;
8045         struct extent_buffer *leaf;
8046         int type;
8047         u32 size;
8048
8049         if (parent > 0)
8050                 type = BTRFS_SHARED_DATA_REF_KEY;
8051         else
8052                 type = BTRFS_EXTENT_DATA_REF_KEY;
8053
8054         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8055
8056         path = btrfs_alloc_path();
8057         if (!path)
8058                 return -ENOMEM;
8059
8060         path->leave_spinning = 1;
8061         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8062                                       ins, size);
8063         if (ret) {
8064                 btrfs_free_path(path);
8065                 return ret;
8066         }
8067
8068         leaf = path->nodes[0];
8069         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8070                                      struct btrfs_extent_item);
8071         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8072         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8073         btrfs_set_extent_flags(leaf, extent_item,
8074                                flags | BTRFS_EXTENT_FLAG_DATA);
8075
8076         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8077         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8078         if (parent > 0) {
8079                 struct btrfs_shared_data_ref *ref;
8080                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8081                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8082                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8083         } else {
8084                 struct btrfs_extent_data_ref *ref;
8085                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8086                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8087                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8088                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8089                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8090         }
8091
8092         btrfs_mark_buffer_dirty(path->nodes[0]);
8093         btrfs_free_path(path);
8094
8095         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8096                                           ins->offset);
8097         if (ret)
8098                 return ret;
8099
8100         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
8101         if (ret) { /* -ENOENT, logic error */
8102                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8103                         ins->objectid, ins->offset);
8104                 BUG();
8105         }
8106         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
8107         return ret;
8108 }
8109
8110 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8111                                      struct btrfs_root *root,
8112                                      u64 parent, u64 root_objectid,
8113                                      u64 flags, struct btrfs_disk_key *key,
8114                                      int level, struct btrfs_key *ins)
8115 {
8116         int ret;
8117         struct btrfs_fs_info *fs_info = root->fs_info;
8118         struct btrfs_extent_item *extent_item;
8119         struct btrfs_tree_block_info *block_info;
8120         struct btrfs_extent_inline_ref *iref;
8121         struct btrfs_path *path;
8122         struct extent_buffer *leaf;
8123         u32 size = sizeof(*extent_item) + sizeof(*iref);
8124         u64 num_bytes = ins->offset;
8125         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8126                                                  SKINNY_METADATA);
8127
8128         if (!skinny_metadata)
8129                 size += sizeof(*block_info);
8130
8131         path = btrfs_alloc_path();
8132         if (!path) {
8133                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8134                                                    root->nodesize);
8135                 return -ENOMEM;
8136         }
8137
8138         path->leave_spinning = 1;
8139         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8140                                       ins, size);
8141         if (ret) {
8142                 btrfs_free_path(path);
8143                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8144                                                    root->nodesize);
8145                 return ret;
8146         }
8147
8148         leaf = path->nodes[0];
8149         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8150                                      struct btrfs_extent_item);
8151         btrfs_set_extent_refs(leaf, extent_item, 1);
8152         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8153         btrfs_set_extent_flags(leaf, extent_item,
8154                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8155
8156         if (skinny_metadata) {
8157                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8158                 num_bytes = root->nodesize;
8159         } else {
8160                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8161                 btrfs_set_tree_block_key(leaf, block_info, key);
8162                 btrfs_set_tree_block_level(leaf, block_info, level);
8163                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8164         }
8165
8166         if (parent > 0) {
8167                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8168                 btrfs_set_extent_inline_ref_type(leaf, iref,
8169                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8170                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8171         } else {
8172                 btrfs_set_extent_inline_ref_type(leaf, iref,
8173                                                  BTRFS_TREE_BLOCK_REF_KEY);
8174                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8175         }
8176
8177         btrfs_mark_buffer_dirty(leaf);
8178         btrfs_free_path(path);
8179
8180         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8181                                           num_bytes);
8182         if (ret)
8183                 return ret;
8184
8185         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
8186                                  1);
8187         if (ret) { /* -ENOENT, logic error */
8188                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8189                         ins->objectid, ins->offset);
8190                 BUG();
8191         }
8192
8193         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
8194         return ret;
8195 }
8196
8197 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8198                                      struct btrfs_root *root,
8199                                      u64 root_objectid, u64 owner,
8200                                      u64 offset, u64 ram_bytes,
8201                                      struct btrfs_key *ins)
8202 {
8203         int ret;
8204
8205         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8206
8207         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
8208                                          ins->offset, 0,
8209                                          root_objectid, owner, offset,
8210                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
8211                                          NULL);
8212         return ret;
8213 }
8214
8215 /*
8216  * this is used by the tree logging recovery code.  It records that
8217  * an extent has been allocated and makes sure to clear the free
8218  * space cache bits as well
8219  */
8220 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8221                                    struct btrfs_root *root,
8222                                    u64 root_objectid, u64 owner, u64 offset,
8223                                    struct btrfs_key *ins)
8224 {
8225         int ret;
8226         struct btrfs_block_group_cache *block_group;
8227         struct btrfs_space_info *space_info;
8228
8229         /*
8230          * Mixed block groups will exclude before processing the log so we only
8231          * need to do the exclude dance if this fs isn't mixed.
8232          */
8233         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
8234                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
8235                 if (ret)
8236                         return ret;
8237         }
8238
8239         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8240         if (!block_group)
8241                 return -EINVAL;
8242
8243         space_info = block_group->space_info;
8244         spin_lock(&space_info->lock);
8245         spin_lock(&block_group->lock);
8246         space_info->bytes_reserved += ins->offset;
8247         block_group->reserved += ins->offset;
8248         spin_unlock(&block_group->lock);
8249         spin_unlock(&space_info->lock);
8250
8251         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8252                                          0, owner, offset, ins, 1);
8253         btrfs_put_block_group(block_group);
8254         return ret;
8255 }
8256
8257 static struct extent_buffer *
8258 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8259                       u64 bytenr, int level)
8260 {
8261         struct extent_buffer *buf;
8262
8263         buf = btrfs_find_create_tree_block(root, bytenr);
8264         if (IS_ERR(buf))
8265                 return buf;
8266
8267         btrfs_set_header_generation(buf, trans->transid);
8268         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8269         btrfs_tree_lock(buf);
8270         clean_tree_block(trans, root->fs_info, buf);
8271         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8272
8273         btrfs_set_lock_blocking(buf);
8274         set_extent_buffer_uptodate(buf);
8275
8276         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8277                 buf->log_index = root->log_transid % 2;
8278                 /*
8279                  * we allow two log transactions at a time, use different
8280                  * EXENT bit to differentiate dirty pages.
8281                  */
8282                 if (buf->log_index == 0)
8283                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8284                                         buf->start + buf->len - 1, GFP_NOFS);
8285                 else
8286                         set_extent_new(&root->dirty_log_pages, buf->start,
8287                                         buf->start + buf->len - 1);
8288         } else {
8289                 buf->log_index = -1;
8290                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8291                          buf->start + buf->len - 1, GFP_NOFS);
8292         }
8293         trans->dirty = true;
8294         /* this returns a buffer locked for blocking */
8295         return buf;
8296 }
8297
8298 static struct btrfs_block_rsv *
8299 use_block_rsv(struct btrfs_trans_handle *trans,
8300               struct btrfs_root *root, u32 blocksize)
8301 {
8302         struct btrfs_block_rsv *block_rsv;
8303         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8304         int ret;
8305         bool global_updated = false;
8306
8307         block_rsv = get_block_rsv(trans, root);
8308
8309         if (unlikely(block_rsv->size == 0))
8310                 goto try_reserve;
8311 again:
8312         ret = block_rsv_use_bytes(block_rsv, blocksize);
8313         if (!ret)
8314                 return block_rsv;
8315
8316         if (block_rsv->failfast)
8317                 return ERR_PTR(ret);
8318
8319         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8320                 global_updated = true;
8321                 update_global_block_rsv(root->fs_info);
8322                 goto again;
8323         }
8324
8325         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
8326                 static DEFINE_RATELIMIT_STATE(_rs,
8327                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8328                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8329                 if (__ratelimit(&_rs))
8330                         WARN(1, KERN_DEBUG
8331                                 "BTRFS: block rsv returned %d\n", ret);
8332         }
8333 try_reserve:
8334         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8335                                      BTRFS_RESERVE_NO_FLUSH);
8336         if (!ret)
8337                 return block_rsv;
8338         /*
8339          * If we couldn't reserve metadata bytes try and use some from
8340          * the global reserve if its space type is the same as the global
8341          * reservation.
8342          */
8343         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8344             block_rsv->space_info == global_rsv->space_info) {
8345                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8346                 if (!ret)
8347                         return global_rsv;
8348         }
8349         return ERR_PTR(ret);
8350 }
8351
8352 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8353                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8354 {
8355         block_rsv_add_bytes(block_rsv, blocksize, 0);
8356         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8357 }
8358
8359 /*
8360  * finds a free extent and does all the dirty work required for allocation
8361  * returns the tree buffer or an ERR_PTR on error.
8362  */
8363 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8364                                         struct btrfs_root *root,
8365                                         u64 parent, u64 root_objectid,
8366                                         struct btrfs_disk_key *key, int level,
8367                                         u64 hint, u64 empty_size)
8368 {
8369         struct btrfs_key ins;
8370         struct btrfs_block_rsv *block_rsv;
8371         struct extent_buffer *buf;
8372         struct btrfs_delayed_extent_op *extent_op;
8373         u64 flags = 0;
8374         int ret;
8375         u32 blocksize = root->nodesize;
8376         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8377                                                  SKINNY_METADATA);
8378
8379 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8380         if (btrfs_is_testing(root->fs_info)) {
8381                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8382                                             level);
8383                 if (!IS_ERR(buf))
8384                         root->alloc_bytenr += blocksize;
8385                 return buf;
8386         }
8387 #endif
8388
8389         block_rsv = use_block_rsv(trans, root, blocksize);
8390         if (IS_ERR(block_rsv))
8391                 return ERR_CAST(block_rsv);
8392
8393         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8394                                    empty_size, hint, &ins, 0, 0);
8395         if (ret)
8396                 goto out_unuse;
8397
8398         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8399         if (IS_ERR(buf)) {
8400                 ret = PTR_ERR(buf);
8401                 goto out_free_reserved;
8402         }
8403
8404         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8405                 if (parent == 0)
8406                         parent = ins.objectid;
8407                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8408         } else
8409                 BUG_ON(parent > 0);
8410
8411         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8412                 extent_op = btrfs_alloc_delayed_extent_op();
8413                 if (!extent_op) {
8414                         ret = -ENOMEM;
8415                         goto out_free_buf;
8416                 }
8417                 if (key)
8418                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8419                 else
8420                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8421                 extent_op->flags_to_set = flags;
8422                 extent_op->update_key = skinny_metadata ? false : true;
8423                 extent_op->update_flags = true;
8424                 extent_op->is_data = false;
8425                 extent_op->level = level;
8426
8427                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8428                                                  ins.objectid, ins.offset,
8429                                                  parent, root_objectid, level,
8430                                                  BTRFS_ADD_DELAYED_EXTENT,
8431                                                  extent_op);
8432                 if (ret)
8433                         goto out_free_delayed;
8434         }
8435         return buf;
8436
8437 out_free_delayed:
8438         btrfs_free_delayed_extent_op(extent_op);
8439 out_free_buf:
8440         free_extent_buffer(buf);
8441 out_free_reserved:
8442         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8443 out_unuse:
8444         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8445         return ERR_PTR(ret);
8446 }
8447
8448 struct walk_control {
8449         u64 refs[BTRFS_MAX_LEVEL];
8450         u64 flags[BTRFS_MAX_LEVEL];
8451         struct btrfs_key update_progress;
8452         int stage;
8453         int level;
8454         int shared_level;
8455         int update_ref;
8456         int keep_locks;
8457         int reada_slot;
8458         int reada_count;
8459         int for_reloc;
8460 };
8461
8462 #define DROP_REFERENCE  1
8463 #define UPDATE_BACKREF  2
8464
8465 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8466                                      struct btrfs_root *root,
8467                                      struct walk_control *wc,
8468                                      struct btrfs_path *path)
8469 {
8470         u64 bytenr;
8471         u64 generation;
8472         u64 refs;
8473         u64 flags;
8474         u32 nritems;
8475         struct btrfs_key key;
8476         struct extent_buffer *eb;
8477         int ret;
8478         int slot;
8479         int nread = 0;
8480
8481         if (path->slots[wc->level] < wc->reada_slot) {
8482                 wc->reada_count = wc->reada_count * 2 / 3;
8483                 wc->reada_count = max(wc->reada_count, 2);
8484         } else {
8485                 wc->reada_count = wc->reada_count * 3 / 2;
8486                 wc->reada_count = min_t(int, wc->reada_count,
8487                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8488         }
8489
8490         eb = path->nodes[wc->level];
8491         nritems = btrfs_header_nritems(eb);
8492
8493         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8494                 if (nread >= wc->reada_count)
8495                         break;
8496
8497                 cond_resched();
8498                 bytenr = btrfs_node_blockptr(eb, slot);
8499                 generation = btrfs_node_ptr_generation(eb, slot);
8500
8501                 if (slot == path->slots[wc->level])
8502                         goto reada;
8503
8504                 if (wc->stage == UPDATE_BACKREF &&
8505                     generation <= root->root_key.offset)
8506                         continue;
8507
8508                 /* We don't lock the tree block, it's OK to be racy here */
8509                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8510                                                wc->level - 1, 1, &refs,
8511                                                &flags);
8512                 /* We don't care about errors in readahead. */
8513                 if (ret < 0)
8514                         continue;
8515                 BUG_ON(refs == 0);
8516
8517                 if (wc->stage == DROP_REFERENCE) {
8518                         if (refs == 1)
8519                                 goto reada;
8520
8521                         if (wc->level == 1 &&
8522                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8523                                 continue;
8524                         if (!wc->update_ref ||
8525                             generation <= root->root_key.offset)
8526                                 continue;
8527                         btrfs_node_key_to_cpu(eb, &key, slot);
8528                         ret = btrfs_comp_cpu_keys(&key,
8529                                                   &wc->update_progress);
8530                         if (ret < 0)
8531                                 continue;
8532                 } else {
8533                         if (wc->level == 1 &&
8534                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8535                                 continue;
8536                 }
8537 reada:
8538                 readahead_tree_block(root, bytenr);
8539                 nread++;
8540         }
8541         wc->reada_slot = slot;
8542 }
8543
8544 static int account_leaf_items(struct btrfs_trans_handle *trans,
8545                               struct btrfs_root *root,
8546                               struct extent_buffer *eb)
8547 {
8548         int nr = btrfs_header_nritems(eb);
8549         int i, extent_type, ret;
8550         struct btrfs_key key;
8551         struct btrfs_file_extent_item *fi;
8552         u64 bytenr, num_bytes;
8553
8554         /* We can be called directly from walk_up_proc() */
8555         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags))
8556                 return 0;
8557
8558         for (i = 0; i < nr; i++) {
8559                 btrfs_item_key_to_cpu(eb, &key, i);
8560
8561                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8562                         continue;
8563
8564                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8565                 /* filter out non qgroup-accountable extents  */
8566                 extent_type = btrfs_file_extent_type(eb, fi);
8567
8568                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8569                         continue;
8570
8571                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8572                 if (!bytenr)
8573                         continue;
8574
8575                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8576
8577                 ret = btrfs_qgroup_insert_dirty_extent(trans, root->fs_info,
8578                                 bytenr, num_bytes, GFP_NOFS);
8579                 if (ret)
8580                         return ret;
8581         }
8582         return 0;
8583 }
8584
8585 /*
8586  * Walk up the tree from the bottom, freeing leaves and any interior
8587  * nodes which have had all slots visited. If a node (leaf or
8588  * interior) is freed, the node above it will have it's slot
8589  * incremented. The root node will never be freed.
8590  *
8591  * At the end of this function, we should have a path which has all
8592  * slots incremented to the next position for a search. If we need to
8593  * read a new node it will be NULL and the node above it will have the
8594  * correct slot selected for a later read.
8595  *
8596  * If we increment the root nodes slot counter past the number of
8597  * elements, 1 is returned to signal completion of the search.
8598  */
8599 static int adjust_slots_upwards(struct btrfs_root *root,
8600                                 struct btrfs_path *path, int root_level)
8601 {
8602         int level = 0;
8603         int nr, slot;
8604         struct extent_buffer *eb;
8605
8606         if (root_level == 0)
8607                 return 1;
8608
8609         while (level <= root_level) {
8610                 eb = path->nodes[level];
8611                 nr = btrfs_header_nritems(eb);
8612                 path->slots[level]++;
8613                 slot = path->slots[level];
8614                 if (slot >= nr || level == 0) {
8615                         /*
8616                          * Don't free the root -  we will detect this
8617                          * condition after our loop and return a
8618                          * positive value for caller to stop walking the tree.
8619                          */
8620                         if (level != root_level) {
8621                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8622                                 path->locks[level] = 0;
8623
8624                                 free_extent_buffer(eb);
8625                                 path->nodes[level] = NULL;
8626                                 path->slots[level] = 0;
8627                         }
8628                 } else {
8629                         /*
8630                          * We have a valid slot to walk back down
8631                          * from. Stop here so caller can process these
8632                          * new nodes.
8633                          */
8634                         break;
8635                 }
8636
8637                 level++;
8638         }
8639
8640         eb = path->nodes[root_level];
8641         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8642                 return 1;
8643
8644         return 0;
8645 }
8646
8647 /*
8648  * root_eb is the subtree root and is locked before this function is called.
8649  */
8650 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8651                                   struct btrfs_root *root,
8652                                   struct extent_buffer *root_eb,
8653                                   u64 root_gen,
8654                                   int root_level)
8655 {
8656         int ret = 0;
8657         int level;
8658         struct extent_buffer *eb = root_eb;
8659         struct btrfs_path *path = NULL;
8660
8661         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8662         BUG_ON(root_eb == NULL);
8663
8664         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags))
8665                 return 0;
8666
8667         if (!extent_buffer_uptodate(root_eb)) {
8668                 ret = btrfs_read_buffer(root_eb, root_gen);
8669                 if (ret)
8670                         goto out;
8671         }
8672
8673         if (root_level == 0) {
8674                 ret = account_leaf_items(trans, root, root_eb);
8675                 goto out;
8676         }
8677
8678         path = btrfs_alloc_path();
8679         if (!path)
8680                 return -ENOMEM;
8681
8682         /*
8683          * Walk down the tree.  Missing extent blocks are filled in as
8684          * we go. Metadata is accounted every time we read a new
8685          * extent block.
8686          *
8687          * When we reach a leaf, we account for file extent items in it,
8688          * walk back up the tree (adjusting slot pointers as we go)
8689          * and restart the search process.
8690          */
8691         extent_buffer_get(root_eb); /* For path */
8692         path->nodes[root_level] = root_eb;
8693         path->slots[root_level] = 0;
8694         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8695 walk_down:
8696         level = root_level;
8697         while (level >= 0) {
8698                 if (path->nodes[level] == NULL) {
8699                         int parent_slot;
8700                         u64 child_gen;
8701                         u64 child_bytenr;
8702
8703                         /* We need to get child blockptr/gen from
8704                          * parent before we can read it. */
8705                         eb = path->nodes[level + 1];
8706                         parent_slot = path->slots[level + 1];
8707                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8708                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8709
8710                         eb = read_tree_block(root, child_bytenr, child_gen);
8711                         if (IS_ERR(eb)) {
8712                                 ret = PTR_ERR(eb);
8713                                 goto out;
8714                         } else if (!extent_buffer_uptodate(eb)) {
8715                                 free_extent_buffer(eb);
8716                                 ret = -EIO;
8717                                 goto out;
8718                         }
8719
8720                         path->nodes[level] = eb;
8721                         path->slots[level] = 0;
8722
8723                         btrfs_tree_read_lock(eb);
8724                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8725                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8726
8727                         ret = btrfs_qgroup_insert_dirty_extent(trans,
8728                                         root->fs_info, child_bytenr,
8729                                         root->nodesize, GFP_NOFS);
8730                         if (ret)
8731                                 goto out;
8732                 }
8733
8734                 if (level == 0) {
8735                         ret = account_leaf_items(trans, root, path->nodes[level]);
8736                         if (ret)
8737                                 goto out;
8738
8739                         /* Nonzero return here means we completed our search */
8740                         ret = adjust_slots_upwards(root, path, root_level);
8741                         if (ret)
8742                                 break;
8743
8744                         /* Restart search with new slots */
8745                         goto walk_down;
8746                 }
8747
8748                 level--;
8749         }
8750
8751         ret = 0;
8752 out:
8753         btrfs_free_path(path);
8754
8755         return ret;
8756 }
8757
8758 /*
8759  * helper to process tree block while walking down the tree.
8760  *
8761  * when wc->stage == UPDATE_BACKREF, this function updates
8762  * back refs for pointers in the block.
8763  *
8764  * NOTE: return value 1 means we should stop walking down.
8765  */
8766 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8767                                    struct btrfs_root *root,
8768                                    struct btrfs_path *path,
8769                                    struct walk_control *wc, int lookup_info)
8770 {
8771         int level = wc->level;
8772         struct extent_buffer *eb = path->nodes[level];
8773         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8774         int ret;
8775
8776         if (wc->stage == UPDATE_BACKREF &&
8777             btrfs_header_owner(eb) != root->root_key.objectid)
8778                 return 1;
8779
8780         /*
8781          * when reference count of tree block is 1, it won't increase
8782          * again. once full backref flag is set, we never clear it.
8783          */
8784         if (lookup_info &&
8785             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8786              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8787                 BUG_ON(!path->locks[level]);
8788                 ret = btrfs_lookup_extent_info(trans, root,
8789                                                eb->start, level, 1,
8790                                                &wc->refs[level],
8791                                                &wc->flags[level]);
8792                 BUG_ON(ret == -ENOMEM);
8793                 if (ret)
8794                         return ret;
8795                 BUG_ON(wc->refs[level] == 0);
8796         }
8797
8798         if (wc->stage == DROP_REFERENCE) {
8799                 if (wc->refs[level] > 1)
8800                         return 1;
8801
8802                 if (path->locks[level] && !wc->keep_locks) {
8803                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8804                         path->locks[level] = 0;
8805                 }
8806                 return 0;
8807         }
8808
8809         /* wc->stage == UPDATE_BACKREF */
8810         if (!(wc->flags[level] & flag)) {
8811                 BUG_ON(!path->locks[level]);
8812                 ret = btrfs_inc_ref(trans, root, eb, 1);
8813                 BUG_ON(ret); /* -ENOMEM */
8814                 ret = btrfs_dec_ref(trans, root, eb, 0);
8815                 BUG_ON(ret); /* -ENOMEM */
8816                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8817                                                   eb->len, flag,
8818                                                   btrfs_header_level(eb), 0);
8819                 BUG_ON(ret); /* -ENOMEM */
8820                 wc->flags[level] |= flag;
8821         }
8822
8823         /*
8824          * the block is shared by multiple trees, so it's not good to
8825          * keep the tree lock
8826          */
8827         if (path->locks[level] && level > 0) {
8828                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8829                 path->locks[level] = 0;
8830         }
8831         return 0;
8832 }
8833
8834 /*
8835  * helper to process tree block pointer.
8836  *
8837  * when wc->stage == DROP_REFERENCE, this function checks
8838  * reference count of the block pointed to. if the block
8839  * is shared and we need update back refs for the subtree
8840  * rooted at the block, this function changes wc->stage to
8841  * UPDATE_BACKREF. if the block is shared and there is no
8842  * need to update back, this function drops the reference
8843  * to the block.
8844  *
8845  * NOTE: return value 1 means we should stop walking down.
8846  */
8847 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8848                                  struct btrfs_root *root,
8849                                  struct btrfs_path *path,
8850                                  struct walk_control *wc, int *lookup_info)
8851 {
8852         u64 bytenr;
8853         u64 generation;
8854         u64 parent;
8855         u32 blocksize;
8856         struct btrfs_key key;
8857         struct extent_buffer *next;
8858         int level = wc->level;
8859         int reada = 0;
8860         int ret = 0;
8861         bool need_account = false;
8862
8863         generation = btrfs_node_ptr_generation(path->nodes[level],
8864                                                path->slots[level]);
8865         /*
8866          * if the lower level block was created before the snapshot
8867          * was created, we know there is no need to update back refs
8868          * for the subtree
8869          */
8870         if (wc->stage == UPDATE_BACKREF &&
8871             generation <= root->root_key.offset) {
8872                 *lookup_info = 1;
8873                 return 1;
8874         }
8875
8876         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8877         blocksize = root->nodesize;
8878
8879         next = btrfs_find_tree_block(root->fs_info, bytenr);
8880         if (!next) {
8881                 next = btrfs_find_create_tree_block(root, bytenr);
8882                 if (IS_ERR(next))
8883                         return PTR_ERR(next);
8884
8885                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8886                                                level - 1);
8887                 reada = 1;
8888         }
8889         btrfs_tree_lock(next);
8890         btrfs_set_lock_blocking(next);
8891
8892         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8893                                        &wc->refs[level - 1],
8894                                        &wc->flags[level - 1]);
8895         if (ret < 0)
8896                 goto out_unlock;
8897
8898         if (unlikely(wc->refs[level - 1] == 0)) {
8899                 btrfs_err(root->fs_info, "Missing references.");
8900                 ret = -EIO;
8901                 goto out_unlock;
8902         }
8903         *lookup_info = 0;
8904
8905         if (wc->stage == DROP_REFERENCE) {
8906                 if (wc->refs[level - 1] > 1) {
8907                         need_account = true;
8908                         if (level == 1 &&
8909                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8910                                 goto skip;
8911
8912                         if (!wc->update_ref ||
8913                             generation <= root->root_key.offset)
8914                                 goto skip;
8915
8916                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8917                                               path->slots[level]);
8918                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8919                         if (ret < 0)
8920                                 goto skip;
8921
8922                         wc->stage = UPDATE_BACKREF;
8923                         wc->shared_level = level - 1;
8924                 }
8925         } else {
8926                 if (level == 1 &&
8927                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8928                         goto skip;
8929         }
8930
8931         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8932                 btrfs_tree_unlock(next);
8933                 free_extent_buffer(next);
8934                 next = NULL;
8935                 *lookup_info = 1;
8936         }
8937
8938         if (!next) {
8939                 if (reada && level == 1)
8940                         reada_walk_down(trans, root, wc, path);
8941                 next = read_tree_block(root, bytenr, generation);
8942                 if (IS_ERR(next)) {
8943                         return PTR_ERR(next);
8944                 } else if (!extent_buffer_uptodate(next)) {
8945                         free_extent_buffer(next);
8946                         return -EIO;
8947                 }
8948                 btrfs_tree_lock(next);
8949                 btrfs_set_lock_blocking(next);
8950         }
8951
8952         level--;
8953         ASSERT(level == btrfs_header_level(next));
8954         if (level != btrfs_header_level(next)) {
8955                 btrfs_err(root->fs_info, "mismatched level");
8956                 ret = -EIO;
8957                 goto out_unlock;
8958         }
8959         path->nodes[level] = next;
8960         path->slots[level] = 0;
8961         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8962         wc->level = level;
8963         if (wc->level == 1)
8964                 wc->reada_slot = 0;
8965         return 0;
8966 skip:
8967         wc->refs[level - 1] = 0;
8968         wc->flags[level - 1] = 0;
8969         if (wc->stage == DROP_REFERENCE) {
8970                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8971                         parent = path->nodes[level]->start;
8972                 } else {
8973                         ASSERT(root->root_key.objectid ==
8974                                btrfs_header_owner(path->nodes[level]));
8975                         if (root->root_key.objectid !=
8976                             btrfs_header_owner(path->nodes[level])) {
8977                                 btrfs_err(root->fs_info,
8978                                                 "mismatched block owner");
8979                                 ret = -EIO;
8980                                 goto out_unlock;
8981                         }
8982                         parent = 0;
8983                 }
8984
8985                 if (need_account) {
8986                         ret = account_shared_subtree(trans, root, next,
8987                                                      generation, level - 1);
8988                         if (ret) {
8989                                 btrfs_err_rl(root->fs_info,
8990                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8991                                              ret);
8992                         }
8993                 }
8994                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8995                                 root->root_key.objectid, level - 1, 0);
8996                 if (ret)
8997                         goto out_unlock;
8998         }
8999
9000         *lookup_info = 1;
9001         ret = 1;
9002
9003 out_unlock:
9004         btrfs_tree_unlock(next);
9005         free_extent_buffer(next);
9006
9007         return ret;
9008 }
9009
9010 /*
9011  * helper to process tree block while walking up the tree.
9012  *
9013  * when wc->stage == DROP_REFERENCE, this function drops
9014  * reference count on the block.
9015  *
9016  * when wc->stage == UPDATE_BACKREF, this function changes
9017  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9018  * to UPDATE_BACKREF previously while processing the block.
9019  *
9020  * NOTE: return value 1 means we should stop walking up.
9021  */
9022 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9023                                  struct btrfs_root *root,
9024                                  struct btrfs_path *path,
9025                                  struct walk_control *wc)
9026 {
9027         int ret;
9028         int level = wc->level;
9029         struct extent_buffer *eb = path->nodes[level];
9030         u64 parent = 0;
9031
9032         if (wc->stage == UPDATE_BACKREF) {
9033                 BUG_ON(wc->shared_level < level);
9034                 if (level < wc->shared_level)
9035                         goto out;
9036
9037                 ret = find_next_key(path, level + 1, &wc->update_progress);
9038                 if (ret > 0)
9039                         wc->update_ref = 0;
9040
9041                 wc->stage = DROP_REFERENCE;
9042                 wc->shared_level = -1;
9043                 path->slots[level] = 0;
9044
9045                 /*
9046                  * check reference count again if the block isn't locked.
9047                  * we should start walking down the tree again if reference
9048                  * count is one.
9049                  */
9050                 if (!path->locks[level]) {
9051                         BUG_ON(level == 0);
9052                         btrfs_tree_lock(eb);
9053                         btrfs_set_lock_blocking(eb);
9054                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9055
9056                         ret = btrfs_lookup_extent_info(trans, root,
9057                                                        eb->start, level, 1,
9058                                                        &wc->refs[level],
9059                                                        &wc->flags[level]);
9060                         if (ret < 0) {
9061                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9062                                 path->locks[level] = 0;
9063                                 return ret;
9064                         }
9065                         BUG_ON(wc->refs[level] == 0);
9066                         if (wc->refs[level] == 1) {
9067                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9068                                 path->locks[level] = 0;
9069                                 return 1;
9070                         }
9071                 }
9072         }
9073
9074         /* wc->stage == DROP_REFERENCE */
9075         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9076
9077         if (wc->refs[level] == 1) {
9078                 if (level == 0) {
9079                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9080                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9081                         else
9082                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9083                         BUG_ON(ret); /* -ENOMEM */
9084                         ret = account_leaf_items(trans, root, eb);
9085                         if (ret) {
9086                                 btrfs_err_rl(root->fs_info,
9087                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9088                                              ret);
9089                         }
9090                 }
9091                 /* make block locked assertion in clean_tree_block happy */
9092                 if (!path->locks[level] &&
9093                     btrfs_header_generation(eb) == trans->transid) {
9094                         btrfs_tree_lock(eb);
9095                         btrfs_set_lock_blocking(eb);
9096                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9097                 }
9098                 clean_tree_block(trans, root->fs_info, eb);
9099         }
9100
9101         if (eb == root->node) {
9102                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9103                         parent = eb->start;
9104                 else
9105                         BUG_ON(root->root_key.objectid !=
9106                                btrfs_header_owner(eb));
9107         } else {
9108                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9109                         parent = path->nodes[level + 1]->start;
9110                 else
9111                         BUG_ON(root->root_key.objectid !=
9112                                btrfs_header_owner(path->nodes[level + 1]));
9113         }
9114
9115         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9116 out:
9117         wc->refs[level] = 0;
9118         wc->flags[level] = 0;
9119         return 0;
9120 }
9121
9122 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9123                                    struct btrfs_root *root,
9124                                    struct btrfs_path *path,
9125                                    struct walk_control *wc)
9126 {
9127         int level = wc->level;
9128         int lookup_info = 1;
9129         int ret;
9130
9131         while (level >= 0) {
9132                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9133                 if (ret > 0)
9134                         break;
9135
9136                 if (level == 0)
9137                         break;
9138
9139                 if (path->slots[level] >=
9140                     btrfs_header_nritems(path->nodes[level]))
9141                         break;
9142
9143                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9144                 if (ret > 0) {
9145                         path->slots[level]++;
9146                         continue;
9147                 } else if (ret < 0)
9148                         return ret;
9149                 level = wc->level;
9150         }
9151         return 0;
9152 }
9153
9154 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9155                                  struct btrfs_root *root,
9156                                  struct btrfs_path *path,
9157                                  struct walk_control *wc, int max_level)
9158 {
9159         int level = wc->level;
9160         int ret;
9161
9162         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9163         while (level < max_level && path->nodes[level]) {
9164                 wc->level = level;
9165                 if (path->slots[level] + 1 <
9166                     btrfs_header_nritems(path->nodes[level])) {
9167                         path->slots[level]++;
9168                         return 0;
9169                 } else {
9170                         ret = walk_up_proc(trans, root, path, wc);
9171                         if (ret > 0)
9172                                 return 0;
9173
9174                         if (path->locks[level]) {
9175                                 btrfs_tree_unlock_rw(path->nodes[level],
9176                                                      path->locks[level]);
9177                                 path->locks[level] = 0;
9178                         }
9179                         free_extent_buffer(path->nodes[level]);
9180                         path->nodes[level] = NULL;
9181                         level++;
9182                 }
9183         }
9184         return 1;
9185 }
9186
9187 /*
9188  * drop a subvolume tree.
9189  *
9190  * this function traverses the tree freeing any blocks that only
9191  * referenced by the tree.
9192  *
9193  * when a shared tree block is found. this function decreases its
9194  * reference count by one. if update_ref is true, this function
9195  * also make sure backrefs for the shared block and all lower level
9196  * blocks are properly updated.
9197  *
9198  * If called with for_reloc == 0, may exit early with -EAGAIN
9199  */
9200 int btrfs_drop_snapshot(struct btrfs_root *root,
9201                          struct btrfs_block_rsv *block_rsv, int update_ref,
9202                          int for_reloc)
9203 {
9204         struct btrfs_fs_info *fs_info = root->fs_info;
9205         struct btrfs_path *path;
9206         struct btrfs_trans_handle *trans;
9207         struct btrfs_root *tree_root = fs_info->tree_root;
9208         struct btrfs_root_item *root_item = &root->root_item;
9209         struct walk_control *wc;
9210         struct btrfs_key key;
9211         int err = 0;
9212         int ret;
9213         int level;
9214         bool root_dropped = false;
9215
9216         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9217
9218         path = btrfs_alloc_path();
9219         if (!path) {
9220                 err = -ENOMEM;
9221                 goto out;
9222         }
9223
9224         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9225         if (!wc) {
9226                 btrfs_free_path(path);
9227                 err = -ENOMEM;
9228                 goto out;
9229         }
9230
9231         trans = btrfs_start_transaction(tree_root, 0);
9232         if (IS_ERR(trans)) {
9233                 err = PTR_ERR(trans);
9234                 goto out_free;
9235         }
9236
9237         if (block_rsv)
9238                 trans->block_rsv = block_rsv;
9239
9240         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9241                 level = btrfs_header_level(root->node);
9242                 path->nodes[level] = btrfs_lock_root_node(root);
9243                 btrfs_set_lock_blocking(path->nodes[level]);
9244                 path->slots[level] = 0;
9245                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9246                 memset(&wc->update_progress, 0,
9247                        sizeof(wc->update_progress));
9248         } else {
9249                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9250                 memcpy(&wc->update_progress, &key,
9251                        sizeof(wc->update_progress));
9252
9253                 level = root_item->drop_level;
9254                 BUG_ON(level == 0);
9255                 path->lowest_level = level;
9256                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9257                 path->lowest_level = 0;
9258                 if (ret < 0) {
9259                         err = ret;
9260                         goto out_end_trans;
9261                 }
9262                 WARN_ON(ret > 0);
9263
9264                 /*
9265                  * unlock our path, this is safe because only this
9266                  * function is allowed to delete this snapshot
9267                  */
9268                 btrfs_unlock_up_safe(path, 0);
9269
9270                 level = btrfs_header_level(root->node);
9271                 while (1) {
9272                         btrfs_tree_lock(path->nodes[level]);
9273                         btrfs_set_lock_blocking(path->nodes[level]);
9274                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9275
9276                         ret = btrfs_lookup_extent_info(trans, root,
9277                                                 path->nodes[level]->start,
9278                                                 level, 1, &wc->refs[level],
9279                                                 &wc->flags[level]);
9280                         if (ret < 0) {
9281                                 err = ret;
9282                                 goto out_end_trans;
9283                         }
9284                         BUG_ON(wc->refs[level] == 0);
9285
9286                         if (level == root_item->drop_level)
9287                                 break;
9288
9289                         btrfs_tree_unlock(path->nodes[level]);
9290                         path->locks[level] = 0;
9291                         WARN_ON(wc->refs[level] != 1);
9292                         level--;
9293                 }
9294         }
9295
9296         wc->level = level;
9297         wc->shared_level = -1;
9298         wc->stage = DROP_REFERENCE;
9299         wc->update_ref = update_ref;
9300         wc->keep_locks = 0;
9301         wc->for_reloc = for_reloc;
9302         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9303
9304         while (1) {
9305
9306                 ret = walk_down_tree(trans, root, path, wc);
9307                 if (ret < 0) {
9308                         err = ret;
9309                         break;
9310                 }
9311
9312                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9313                 if (ret < 0) {
9314                         err = ret;
9315                         break;
9316                 }
9317
9318                 if (ret > 0) {
9319                         BUG_ON(wc->stage != DROP_REFERENCE);
9320                         break;
9321                 }
9322
9323                 if (wc->stage == DROP_REFERENCE) {
9324                         level = wc->level;
9325                         btrfs_node_key(path->nodes[level],
9326                                        &root_item->drop_progress,
9327                                        path->slots[level]);
9328                         root_item->drop_level = level;
9329                 }
9330
9331                 BUG_ON(wc->level == 0);
9332                 if (btrfs_should_end_transaction(trans, tree_root) ||
9333                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9334                         ret = btrfs_update_root(trans, tree_root,
9335                                                 &root->root_key,
9336                                                 root_item);
9337                         if (ret) {
9338                                 btrfs_abort_transaction(trans, ret);
9339                                 err = ret;
9340                                 goto out_end_trans;
9341                         }
9342
9343                         btrfs_end_transaction_throttle(trans, tree_root);
9344                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9345                                 btrfs_debug(fs_info,
9346                                             "drop snapshot early exit");
9347                                 err = -EAGAIN;
9348                                 goto out_free;
9349                         }
9350
9351                         trans = btrfs_start_transaction(tree_root, 0);
9352                         if (IS_ERR(trans)) {
9353                                 err = PTR_ERR(trans);
9354                                 goto out_free;
9355                         }
9356                         if (block_rsv)
9357                                 trans->block_rsv = block_rsv;
9358                 }
9359         }
9360         btrfs_release_path(path);
9361         if (err)
9362                 goto out_end_trans;
9363
9364         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9365         if (ret) {
9366                 btrfs_abort_transaction(trans, ret);
9367                 goto out_end_trans;
9368         }
9369
9370         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9371                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9372                                       NULL, NULL);
9373                 if (ret < 0) {
9374                         btrfs_abort_transaction(trans, ret);
9375                         err = ret;
9376                         goto out_end_trans;
9377                 } else if (ret > 0) {
9378                         /* if we fail to delete the orphan item this time
9379                          * around, it'll get picked up the next time.
9380                          *
9381                          * The most common failure here is just -ENOENT.
9382                          */
9383                         btrfs_del_orphan_item(trans, tree_root,
9384                                               root->root_key.objectid);
9385                 }
9386         }
9387
9388         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9389                 btrfs_add_dropped_root(trans, root);
9390         } else {
9391                 free_extent_buffer(root->node);
9392                 free_extent_buffer(root->commit_root);
9393                 btrfs_put_fs_root(root);
9394         }
9395         root_dropped = true;
9396 out_end_trans:
9397         btrfs_end_transaction_throttle(trans, tree_root);
9398 out_free:
9399         kfree(wc);
9400         btrfs_free_path(path);
9401 out:
9402         /*
9403          * So if we need to stop dropping the snapshot for whatever reason we
9404          * need to make sure to add it back to the dead root list so that we
9405          * keep trying to do the work later.  This also cleans up roots if we
9406          * don't have it in the radix (like when we recover after a power fail
9407          * or unmount) so we don't leak memory.
9408          */
9409         if (!for_reloc && root_dropped == false)
9410                 btrfs_add_dead_root(root);
9411         if (err && err != -EAGAIN)
9412                 btrfs_handle_fs_error(fs_info, err, NULL);
9413         return err;
9414 }
9415
9416 /*
9417  * drop subtree rooted at tree block 'node'.
9418  *
9419  * NOTE: this function will unlock and release tree block 'node'
9420  * only used by relocation code
9421  */
9422 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9423                         struct btrfs_root *root,
9424                         struct extent_buffer *node,
9425                         struct extent_buffer *parent)
9426 {
9427         struct btrfs_path *path;
9428         struct walk_control *wc;
9429         int level;
9430         int parent_level;
9431         int ret = 0;
9432         int wret;
9433
9434         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9435
9436         path = btrfs_alloc_path();
9437         if (!path)
9438                 return -ENOMEM;
9439
9440         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9441         if (!wc) {
9442                 btrfs_free_path(path);
9443                 return -ENOMEM;
9444         }
9445
9446         btrfs_assert_tree_locked(parent);
9447         parent_level = btrfs_header_level(parent);
9448         extent_buffer_get(parent);
9449         path->nodes[parent_level] = parent;
9450         path->slots[parent_level] = btrfs_header_nritems(parent);
9451
9452         btrfs_assert_tree_locked(node);
9453         level = btrfs_header_level(node);
9454         path->nodes[level] = node;
9455         path->slots[level] = 0;
9456         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9457
9458         wc->refs[parent_level] = 1;
9459         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9460         wc->level = level;
9461         wc->shared_level = -1;
9462         wc->stage = DROP_REFERENCE;
9463         wc->update_ref = 0;
9464         wc->keep_locks = 1;
9465         wc->for_reloc = 1;
9466         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9467
9468         while (1) {
9469                 wret = walk_down_tree(trans, root, path, wc);
9470                 if (wret < 0) {
9471                         ret = wret;
9472                         break;
9473                 }
9474
9475                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9476                 if (wret < 0)
9477                         ret = wret;
9478                 if (wret != 0)
9479                         break;
9480         }
9481
9482         kfree(wc);
9483         btrfs_free_path(path);
9484         return ret;
9485 }
9486
9487 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9488 {
9489         u64 num_devices;
9490         u64 stripped;
9491
9492         /*
9493          * if restripe for this chunk_type is on pick target profile and
9494          * return, otherwise do the usual balance
9495          */
9496         stripped = get_restripe_target(root->fs_info, flags);
9497         if (stripped)
9498                 return extended_to_chunk(stripped);
9499
9500         num_devices = root->fs_info->fs_devices->rw_devices;
9501
9502         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9503                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9504                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9505
9506         if (num_devices == 1) {
9507                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9508                 stripped = flags & ~stripped;
9509
9510                 /* turn raid0 into single device chunks */
9511                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9512                         return stripped;
9513
9514                 /* turn mirroring into duplication */
9515                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9516                              BTRFS_BLOCK_GROUP_RAID10))
9517                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9518         } else {
9519                 /* they already had raid on here, just return */
9520                 if (flags & stripped)
9521                         return flags;
9522
9523                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9524                 stripped = flags & ~stripped;
9525
9526                 /* switch duplicated blocks with raid1 */
9527                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9528                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9529
9530                 /* this is drive concat, leave it alone */
9531         }
9532
9533         return flags;
9534 }
9535
9536 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9537 {
9538         struct btrfs_space_info *sinfo = cache->space_info;
9539         u64 num_bytes;
9540         u64 min_allocable_bytes;
9541         int ret = -ENOSPC;
9542
9543         /*
9544          * We need some metadata space and system metadata space for
9545          * allocating chunks in some corner cases until we force to set
9546          * it to be readonly.
9547          */
9548         if ((sinfo->flags &
9549              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9550             !force)
9551                 min_allocable_bytes = SZ_1M;
9552         else
9553                 min_allocable_bytes = 0;
9554
9555         spin_lock(&sinfo->lock);
9556         spin_lock(&cache->lock);
9557
9558         if (cache->ro) {
9559                 cache->ro++;
9560                 ret = 0;
9561                 goto out;
9562         }
9563
9564         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9565                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9566
9567         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9568             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9569             min_allocable_bytes <= sinfo->total_bytes) {
9570                 sinfo->bytes_readonly += num_bytes;
9571                 cache->ro++;
9572                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9573                 ret = 0;
9574         }
9575 out:
9576         spin_unlock(&cache->lock);
9577         spin_unlock(&sinfo->lock);
9578         return ret;
9579 }
9580
9581 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9582                              struct btrfs_block_group_cache *cache)
9583
9584 {
9585         struct btrfs_trans_handle *trans;
9586         u64 alloc_flags;
9587         int ret;
9588
9589 again:
9590         trans = btrfs_join_transaction(root);
9591         if (IS_ERR(trans))
9592                 return PTR_ERR(trans);
9593
9594         /*
9595          * we're not allowed to set block groups readonly after the dirty
9596          * block groups cache has started writing.  If it already started,
9597          * back off and let this transaction commit
9598          */
9599         mutex_lock(&root->fs_info->ro_block_group_mutex);
9600         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9601                 u64 transid = trans->transid;
9602
9603                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9604                 btrfs_end_transaction(trans, root);
9605
9606                 ret = btrfs_wait_for_commit(root, transid);
9607                 if (ret)
9608                         return ret;
9609                 goto again;
9610         }
9611
9612         /*
9613          * if we are changing raid levels, try to allocate a corresponding
9614          * block group with the new raid level.
9615          */
9616         alloc_flags = update_block_group_flags(root, cache->flags);
9617         if (alloc_flags != cache->flags) {
9618                 ret = do_chunk_alloc(trans, root, alloc_flags,
9619                                      CHUNK_ALLOC_FORCE);
9620                 /*
9621                  * ENOSPC is allowed here, we may have enough space
9622                  * already allocated at the new raid level to
9623                  * carry on
9624                  */
9625                 if (ret == -ENOSPC)
9626                         ret = 0;
9627                 if (ret < 0)
9628                         goto out;
9629         }
9630
9631         ret = inc_block_group_ro(cache, 0);
9632         if (!ret)
9633                 goto out;
9634         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9635         ret = do_chunk_alloc(trans, root, alloc_flags,
9636                              CHUNK_ALLOC_FORCE);
9637         if (ret < 0)
9638                 goto out;
9639         ret = inc_block_group_ro(cache, 0);
9640 out:
9641         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9642                 alloc_flags = update_block_group_flags(root, cache->flags);
9643                 lock_chunks(root->fs_info->chunk_root);
9644                 check_system_chunk(trans, root, alloc_flags);
9645                 unlock_chunks(root->fs_info->chunk_root);
9646         }
9647         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9648
9649         btrfs_end_transaction(trans, root);
9650         return ret;
9651 }
9652
9653 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9654                             struct btrfs_root *root, u64 type)
9655 {
9656         u64 alloc_flags = get_alloc_profile(root, type);
9657         return do_chunk_alloc(trans, root, alloc_flags,
9658                               CHUNK_ALLOC_FORCE);
9659 }
9660
9661 /*
9662  * helper to account the unused space of all the readonly block group in the
9663  * space_info. takes mirrors into account.
9664  */
9665 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9666 {
9667         struct btrfs_block_group_cache *block_group;
9668         u64 free_bytes = 0;
9669         int factor;
9670
9671         /* It's df, we don't care if it's racy */
9672         if (list_empty(&sinfo->ro_bgs))
9673                 return 0;
9674
9675         spin_lock(&sinfo->lock);
9676         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9677                 spin_lock(&block_group->lock);
9678
9679                 if (!block_group->ro) {
9680                         spin_unlock(&block_group->lock);
9681                         continue;
9682                 }
9683
9684                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9685                                           BTRFS_BLOCK_GROUP_RAID10 |
9686                                           BTRFS_BLOCK_GROUP_DUP))
9687                         factor = 2;
9688                 else
9689                         factor = 1;
9690
9691                 free_bytes += (block_group->key.offset -
9692                                btrfs_block_group_used(&block_group->item)) *
9693                                factor;
9694
9695                 spin_unlock(&block_group->lock);
9696         }
9697         spin_unlock(&sinfo->lock);
9698
9699         return free_bytes;
9700 }
9701
9702 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9703                               struct btrfs_block_group_cache *cache)
9704 {
9705         struct btrfs_space_info *sinfo = cache->space_info;
9706         u64 num_bytes;
9707
9708         BUG_ON(!cache->ro);
9709
9710         spin_lock(&sinfo->lock);
9711         spin_lock(&cache->lock);
9712         if (!--cache->ro) {
9713                 num_bytes = cache->key.offset - cache->reserved -
9714                             cache->pinned - cache->bytes_super -
9715                             btrfs_block_group_used(&cache->item);
9716                 sinfo->bytes_readonly -= num_bytes;
9717                 list_del_init(&cache->ro_list);
9718         }
9719         spin_unlock(&cache->lock);
9720         spin_unlock(&sinfo->lock);
9721 }
9722
9723 /*
9724  * checks to see if its even possible to relocate this block group.
9725  *
9726  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9727  * ok to go ahead and try.
9728  */
9729 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9730 {
9731         struct btrfs_block_group_cache *block_group;
9732         struct btrfs_space_info *space_info;
9733         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9734         struct btrfs_device *device;
9735         struct btrfs_trans_handle *trans;
9736         u64 min_free;
9737         u64 dev_min = 1;
9738         u64 dev_nr = 0;
9739         u64 target;
9740         int debug;
9741         int index;
9742         int full = 0;
9743         int ret = 0;
9744
9745         debug = btrfs_test_opt(root->fs_info, ENOSPC_DEBUG);
9746
9747         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9748
9749         /* odd, couldn't find the block group, leave it alone */
9750         if (!block_group) {
9751                 if (debug)
9752                         btrfs_warn(root->fs_info,
9753                                    "can't find block group for bytenr %llu",
9754                                    bytenr);
9755                 return -1;
9756         }
9757
9758         min_free = btrfs_block_group_used(&block_group->item);
9759
9760         /* no bytes used, we're good */
9761         if (!min_free)
9762                 goto out;
9763
9764         space_info = block_group->space_info;
9765         spin_lock(&space_info->lock);
9766
9767         full = space_info->full;
9768
9769         /*
9770          * if this is the last block group we have in this space, we can't
9771          * relocate it unless we're able to allocate a new chunk below.
9772          *
9773          * Otherwise, we need to make sure we have room in the space to handle
9774          * all of the extents from this block group.  If we can, we're good
9775          */
9776         if ((space_info->total_bytes != block_group->key.offset) &&
9777             (space_info->bytes_used + space_info->bytes_reserved +
9778              space_info->bytes_pinned + space_info->bytes_readonly +
9779              min_free < space_info->total_bytes)) {
9780                 spin_unlock(&space_info->lock);
9781                 goto out;
9782         }
9783         spin_unlock(&space_info->lock);
9784
9785         /*
9786          * ok we don't have enough space, but maybe we have free space on our
9787          * devices to allocate new chunks for relocation, so loop through our
9788          * alloc devices and guess if we have enough space.  if this block
9789          * group is going to be restriped, run checks against the target
9790          * profile instead of the current one.
9791          */
9792         ret = -1;
9793
9794         /*
9795          * index:
9796          *      0: raid10
9797          *      1: raid1
9798          *      2: dup
9799          *      3: raid0
9800          *      4: single
9801          */
9802         target = get_restripe_target(root->fs_info, block_group->flags);
9803         if (target) {
9804                 index = __get_raid_index(extended_to_chunk(target));
9805         } else {
9806                 /*
9807                  * this is just a balance, so if we were marked as full
9808                  * we know there is no space for a new chunk
9809                  */
9810                 if (full) {
9811                         if (debug)
9812                                 btrfs_warn(root->fs_info,
9813                                         "no space to alloc new chunk for block group %llu",
9814                                         block_group->key.objectid);
9815                         goto out;
9816                 }
9817
9818                 index = get_block_group_index(block_group);
9819         }
9820
9821         if (index == BTRFS_RAID_RAID10) {
9822                 dev_min = 4;
9823                 /* Divide by 2 */
9824                 min_free >>= 1;
9825         } else if (index == BTRFS_RAID_RAID1) {
9826                 dev_min = 2;
9827         } else if (index == BTRFS_RAID_DUP) {
9828                 /* Multiply by 2 */
9829                 min_free <<= 1;
9830         } else if (index == BTRFS_RAID_RAID0) {
9831                 dev_min = fs_devices->rw_devices;
9832                 min_free = div64_u64(min_free, dev_min);
9833         }
9834
9835         /* We need to do this so that we can look at pending chunks */
9836         trans = btrfs_join_transaction(root);
9837         if (IS_ERR(trans)) {
9838                 ret = PTR_ERR(trans);
9839                 goto out;
9840         }
9841
9842         mutex_lock(&root->fs_info->chunk_mutex);
9843         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9844                 u64 dev_offset;
9845
9846                 /*
9847                  * check to make sure we can actually find a chunk with enough
9848                  * space to fit our block group in.
9849                  */
9850                 if (device->total_bytes > device->bytes_used + min_free &&
9851                     !device->is_tgtdev_for_dev_replace) {
9852                         ret = find_free_dev_extent(trans, device, min_free,
9853                                                    &dev_offset, NULL);
9854                         if (!ret)
9855                                 dev_nr++;
9856
9857                         if (dev_nr >= dev_min)
9858                                 break;
9859
9860                         ret = -1;
9861                 }
9862         }
9863         if (debug && ret == -1)
9864                 btrfs_warn(root->fs_info,
9865                         "no space to allocate a new chunk for block group %llu",
9866                         block_group->key.objectid);
9867         mutex_unlock(&root->fs_info->chunk_mutex);
9868         btrfs_end_transaction(trans, root);
9869 out:
9870         btrfs_put_block_group(block_group);
9871         return ret;
9872 }
9873
9874 static int find_first_block_group(struct btrfs_root *root,
9875                 struct btrfs_path *path, struct btrfs_key *key)
9876 {
9877         int ret = 0;
9878         struct btrfs_key found_key;
9879         struct extent_buffer *leaf;
9880         int slot;
9881
9882         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9883         if (ret < 0)
9884                 goto out;
9885
9886         while (1) {
9887                 slot = path->slots[0];
9888                 leaf = path->nodes[0];
9889                 if (slot >= btrfs_header_nritems(leaf)) {
9890                         ret = btrfs_next_leaf(root, path);
9891                         if (ret == 0)
9892                                 continue;
9893                         if (ret < 0)
9894                                 goto out;
9895                         break;
9896                 }
9897                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9898
9899                 if (found_key.objectid >= key->objectid &&
9900                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9901                         struct extent_map_tree *em_tree;
9902                         struct extent_map *em;
9903
9904                         em_tree = &root->fs_info->mapping_tree.map_tree;
9905                         read_lock(&em_tree->lock);
9906                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9907                                                    found_key.offset);
9908                         read_unlock(&em_tree->lock);
9909                         if (!em) {
9910                                 btrfs_err(root->fs_info,
9911                         "logical %llu len %llu found bg but no related chunk",
9912                                           found_key.objectid, found_key.offset);
9913                                 ret = -ENOENT;
9914                         } else {
9915                                 ret = 0;
9916                         }
9917                         free_extent_map(em);
9918                         goto out;
9919                 }
9920                 path->slots[0]++;
9921         }
9922 out:
9923         return ret;
9924 }
9925
9926 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9927 {
9928         struct btrfs_block_group_cache *block_group;
9929         u64 last = 0;
9930
9931         while (1) {
9932                 struct inode *inode;
9933
9934                 block_group = btrfs_lookup_first_block_group(info, last);
9935                 while (block_group) {
9936                         spin_lock(&block_group->lock);
9937                         if (block_group->iref)
9938                                 break;
9939                         spin_unlock(&block_group->lock);
9940                         block_group = next_block_group(info->tree_root,
9941                                                        block_group);
9942                 }
9943                 if (!block_group) {
9944                         if (last == 0)
9945                                 break;
9946                         last = 0;
9947                         continue;
9948                 }
9949
9950                 inode = block_group->inode;
9951                 block_group->iref = 0;
9952                 block_group->inode = NULL;
9953                 spin_unlock(&block_group->lock);
9954                 ASSERT(block_group->io_ctl.inode == NULL);
9955                 iput(inode);
9956                 last = block_group->key.objectid + block_group->key.offset;
9957                 btrfs_put_block_group(block_group);
9958         }
9959 }
9960
9961 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9962 {
9963         struct btrfs_block_group_cache *block_group;
9964         struct btrfs_space_info *space_info;
9965         struct btrfs_caching_control *caching_ctl;
9966         struct rb_node *n;
9967
9968         down_write(&info->commit_root_sem);
9969         while (!list_empty(&info->caching_block_groups)) {
9970                 caching_ctl = list_entry(info->caching_block_groups.next,
9971                                          struct btrfs_caching_control, list);
9972                 list_del(&caching_ctl->list);
9973                 put_caching_control(caching_ctl);
9974         }
9975         up_write(&info->commit_root_sem);
9976
9977         spin_lock(&info->unused_bgs_lock);
9978         while (!list_empty(&info->unused_bgs)) {
9979                 block_group = list_first_entry(&info->unused_bgs,
9980                                                struct btrfs_block_group_cache,
9981                                                bg_list);
9982                 list_del_init(&block_group->bg_list);
9983                 btrfs_put_block_group(block_group);
9984         }
9985         spin_unlock(&info->unused_bgs_lock);
9986
9987         spin_lock(&info->block_group_cache_lock);
9988         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9989                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9990                                        cache_node);
9991                 rb_erase(&block_group->cache_node,
9992                          &info->block_group_cache_tree);
9993                 RB_CLEAR_NODE(&block_group->cache_node);
9994                 spin_unlock(&info->block_group_cache_lock);
9995
9996                 down_write(&block_group->space_info->groups_sem);
9997                 list_del(&block_group->list);
9998                 up_write(&block_group->space_info->groups_sem);
9999
10000                 if (block_group->cached == BTRFS_CACHE_STARTED)
10001                         wait_block_group_cache_done(block_group);
10002
10003                 /*
10004                  * We haven't cached this block group, which means we could
10005                  * possibly have excluded extents on this block group.
10006                  */
10007                 if (block_group->cached == BTRFS_CACHE_NO ||
10008                     block_group->cached == BTRFS_CACHE_ERROR)
10009                         free_excluded_extents(info->extent_root, block_group);
10010
10011                 btrfs_remove_free_space_cache(block_group);
10012                 ASSERT(list_empty(&block_group->dirty_list));
10013                 ASSERT(list_empty(&block_group->io_list));
10014                 ASSERT(list_empty(&block_group->bg_list));
10015                 ASSERT(atomic_read(&block_group->count) == 1);
10016                 btrfs_put_block_group(block_group);
10017
10018                 spin_lock(&info->block_group_cache_lock);
10019         }
10020         spin_unlock(&info->block_group_cache_lock);
10021
10022         /* now that all the block groups are freed, go through and
10023          * free all the space_info structs.  This is only called during
10024          * the final stages of unmount, and so we know nobody is
10025          * using them.  We call synchronize_rcu() once before we start,
10026          * just to be on the safe side.
10027          */
10028         synchronize_rcu();
10029
10030         release_global_block_rsv(info);
10031
10032         while (!list_empty(&info->space_info)) {
10033                 int i;
10034
10035                 space_info = list_entry(info->space_info.next,
10036                                         struct btrfs_space_info,
10037                                         list);
10038
10039                 /*
10040                  * Do not hide this behind enospc_debug, this is actually
10041                  * important and indicates a real bug if this happens.
10042                  */
10043                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10044                             space_info->bytes_reserved > 0 ||
10045                             space_info->bytes_may_use > 0))
10046                         dump_space_info(info, space_info, 0, 0);
10047                 list_del(&space_info->list);
10048                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10049                         struct kobject *kobj;
10050                         kobj = space_info->block_group_kobjs[i];
10051                         space_info->block_group_kobjs[i] = NULL;
10052                         if (kobj) {
10053                                 kobject_del(kobj);
10054                                 kobject_put(kobj);
10055                         }
10056                 }
10057                 kobject_del(&space_info->kobj);
10058                 kobject_put(&space_info->kobj);
10059         }
10060         return 0;
10061 }
10062
10063 static void __link_block_group(struct btrfs_space_info *space_info,
10064                                struct btrfs_block_group_cache *cache)
10065 {
10066         int index = get_block_group_index(cache);
10067         bool first = false;
10068
10069         down_write(&space_info->groups_sem);
10070         if (list_empty(&space_info->block_groups[index]))
10071                 first = true;
10072         list_add_tail(&cache->list, &space_info->block_groups[index]);
10073         up_write(&space_info->groups_sem);
10074
10075         if (first) {
10076                 struct raid_kobject *rkobj;
10077                 int ret;
10078
10079                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10080                 if (!rkobj)
10081                         goto out_err;
10082                 rkobj->raid_type = index;
10083                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10084                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10085                                   "%s", get_raid_name(index));
10086                 if (ret) {
10087                         kobject_put(&rkobj->kobj);
10088                         goto out_err;
10089                 }
10090                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10091         }
10092
10093         return;
10094 out_err:
10095         btrfs_warn(cache->fs_info,
10096                    "failed to add kobject for block cache, ignoring");
10097 }
10098
10099 static struct btrfs_block_group_cache *
10100 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
10101 {
10102         struct btrfs_block_group_cache *cache;
10103
10104         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10105         if (!cache)
10106                 return NULL;
10107
10108         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10109                                         GFP_NOFS);
10110         if (!cache->free_space_ctl) {
10111                 kfree(cache);
10112                 return NULL;
10113         }
10114
10115         cache->key.objectid = start;
10116         cache->key.offset = size;
10117         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10118
10119         cache->sectorsize = root->sectorsize;
10120         cache->fs_info = root->fs_info;
10121         cache->full_stripe_len = btrfs_full_stripe_len(root,
10122                                                &root->fs_info->mapping_tree,
10123                                                start);
10124         set_free_space_tree_thresholds(cache);
10125
10126         atomic_set(&cache->count, 1);
10127         spin_lock_init(&cache->lock);
10128         init_rwsem(&cache->data_rwsem);
10129         INIT_LIST_HEAD(&cache->list);
10130         INIT_LIST_HEAD(&cache->cluster_list);
10131         INIT_LIST_HEAD(&cache->bg_list);
10132         INIT_LIST_HEAD(&cache->ro_list);
10133         INIT_LIST_HEAD(&cache->dirty_list);
10134         INIT_LIST_HEAD(&cache->io_list);
10135         btrfs_init_free_space_ctl(cache);
10136         atomic_set(&cache->trimming, 0);
10137         mutex_init(&cache->free_space_lock);
10138
10139         return cache;
10140 }
10141
10142 int btrfs_read_block_groups(struct btrfs_root *root)
10143 {
10144         struct btrfs_path *path;
10145         int ret;
10146         struct btrfs_block_group_cache *cache;
10147         struct btrfs_fs_info *info = root->fs_info;
10148         struct btrfs_space_info *space_info;
10149         struct btrfs_key key;
10150         struct btrfs_key found_key;
10151         struct extent_buffer *leaf;
10152         int need_clear = 0;
10153         u64 cache_gen;
10154         u64 feature;
10155         int mixed;
10156
10157         feature = btrfs_super_incompat_flags(info->super_copy);
10158         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10159
10160         root = info->extent_root;
10161         key.objectid = 0;
10162         key.offset = 0;
10163         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10164         path = btrfs_alloc_path();
10165         if (!path)
10166                 return -ENOMEM;
10167         path->reada = READA_FORWARD;
10168
10169         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
10170         if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
10171             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
10172                 need_clear = 1;
10173         if (btrfs_test_opt(root->fs_info, CLEAR_CACHE))
10174                 need_clear = 1;
10175
10176         while (1) {
10177                 ret = find_first_block_group(root, path, &key);
10178                 if (ret > 0)
10179                         break;
10180                 if (ret != 0)
10181                         goto error;
10182
10183                 leaf = path->nodes[0];
10184                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10185
10186                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
10187                                                        found_key.offset);
10188                 if (!cache) {
10189                         ret = -ENOMEM;
10190                         goto error;
10191                 }
10192
10193                 if (need_clear) {
10194                         /*
10195                          * When we mount with old space cache, we need to
10196                          * set BTRFS_DC_CLEAR and set dirty flag.
10197                          *
10198                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10199                          *    truncate the old free space cache inode and
10200                          *    setup a new one.
10201                          * b) Setting 'dirty flag' makes sure that we flush
10202                          *    the new space cache info onto disk.
10203                          */
10204                         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
10205                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10206                 }
10207
10208                 read_extent_buffer(leaf, &cache->item,
10209                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10210                                    sizeof(cache->item));
10211                 cache->flags = btrfs_block_group_flags(&cache->item);
10212                 if (!mixed &&
10213                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10214                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10215                         btrfs_err(info,
10216 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10217                                   cache->key.objectid);
10218                         ret = -EINVAL;
10219                         goto error;
10220                 }
10221
10222                 key.objectid = found_key.objectid + found_key.offset;
10223                 btrfs_release_path(path);
10224
10225                 /*
10226                  * We need to exclude the super stripes now so that the space
10227                  * info has super bytes accounted for, otherwise we'll think
10228                  * we have more space than we actually do.
10229                  */
10230                 ret = exclude_super_stripes(root, cache);
10231                 if (ret) {
10232                         /*
10233                          * We may have excluded something, so call this just in
10234                          * case.
10235                          */
10236                         free_excluded_extents(root, cache);
10237                         btrfs_put_block_group(cache);
10238                         goto error;
10239                 }
10240
10241                 /*
10242                  * check for two cases, either we are full, and therefore
10243                  * don't need to bother with the caching work since we won't
10244                  * find any space, or we are empty, and we can just add all
10245                  * the space in and be done with it.  This saves us _alot_ of
10246                  * time, particularly in the full case.
10247                  */
10248                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10249                         cache->last_byte_to_unpin = (u64)-1;
10250                         cache->cached = BTRFS_CACHE_FINISHED;
10251                         free_excluded_extents(root, cache);
10252                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10253                         cache->last_byte_to_unpin = (u64)-1;
10254                         cache->cached = BTRFS_CACHE_FINISHED;
10255                         add_new_free_space(cache, root->fs_info,
10256                                            found_key.objectid,
10257                                            found_key.objectid +
10258                                            found_key.offset);
10259                         free_excluded_extents(root, cache);
10260                 }
10261
10262                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
10263                 if (ret) {
10264                         btrfs_remove_free_space_cache(cache);
10265                         btrfs_put_block_group(cache);
10266                         goto error;
10267                 }
10268
10269                 trace_btrfs_add_block_group(root->fs_info, cache, 0);
10270                 ret = update_space_info(info, cache->flags, found_key.offset,
10271                                         btrfs_block_group_used(&cache->item),
10272                                         cache->bytes_super, &space_info);
10273                 if (ret) {
10274                         btrfs_remove_free_space_cache(cache);
10275                         spin_lock(&info->block_group_cache_lock);
10276                         rb_erase(&cache->cache_node,
10277                                  &info->block_group_cache_tree);
10278                         RB_CLEAR_NODE(&cache->cache_node);
10279                         spin_unlock(&info->block_group_cache_lock);
10280                         btrfs_put_block_group(cache);
10281                         goto error;
10282                 }
10283
10284                 cache->space_info = space_info;
10285
10286                 __link_block_group(space_info, cache);
10287
10288                 set_avail_alloc_bits(root->fs_info, cache->flags);
10289                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10290                         inc_block_group_ro(cache, 1);
10291                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10292                         spin_lock(&info->unused_bgs_lock);
10293                         /* Should always be true but just in case. */
10294                         if (list_empty(&cache->bg_list)) {
10295                                 btrfs_get_block_group(cache);
10296                                 list_add_tail(&cache->bg_list,
10297                                               &info->unused_bgs);
10298                         }
10299                         spin_unlock(&info->unused_bgs_lock);
10300                 }
10301         }
10302
10303         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10304                 if (!(get_alloc_profile(root, space_info->flags) &
10305                       (BTRFS_BLOCK_GROUP_RAID10 |
10306                        BTRFS_BLOCK_GROUP_RAID1 |
10307                        BTRFS_BLOCK_GROUP_RAID5 |
10308                        BTRFS_BLOCK_GROUP_RAID6 |
10309                        BTRFS_BLOCK_GROUP_DUP)))
10310                         continue;
10311                 /*
10312                  * avoid allocating from un-mirrored block group if there are
10313                  * mirrored block groups.
10314                  */
10315                 list_for_each_entry(cache,
10316                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10317                                 list)
10318                         inc_block_group_ro(cache, 1);
10319                 list_for_each_entry(cache,
10320                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10321                                 list)
10322                         inc_block_group_ro(cache, 1);
10323         }
10324
10325         init_global_block_rsv(info);
10326         ret = 0;
10327 error:
10328         btrfs_free_path(path);
10329         return ret;
10330 }
10331
10332 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10333                                        struct btrfs_root *root)
10334 {
10335         struct btrfs_block_group_cache *block_group, *tmp;
10336         struct btrfs_root *extent_root = root->fs_info->extent_root;
10337         struct btrfs_block_group_item item;
10338         struct btrfs_key key;
10339         int ret = 0;
10340         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10341
10342         trans->can_flush_pending_bgs = false;
10343         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10344                 if (ret)
10345                         goto next;
10346
10347                 spin_lock(&block_group->lock);
10348                 memcpy(&item, &block_group->item, sizeof(item));
10349                 memcpy(&key, &block_group->key, sizeof(key));
10350                 spin_unlock(&block_group->lock);
10351
10352                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10353                                         sizeof(item));
10354                 if (ret)
10355                         btrfs_abort_transaction(trans, ret);
10356                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10357                                                key.objectid, key.offset);
10358                 if (ret)
10359                         btrfs_abort_transaction(trans, ret);
10360                 add_block_group_free_space(trans, root->fs_info, block_group);
10361                 /* already aborted the transaction if it failed. */
10362 next:
10363                 list_del_init(&block_group->bg_list);
10364         }
10365         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10366 }
10367
10368 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10369                            struct btrfs_root *root, u64 bytes_used,
10370                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10371                            u64 size)
10372 {
10373         int ret;
10374         struct btrfs_root *extent_root;
10375         struct btrfs_block_group_cache *cache;
10376         extent_root = root->fs_info->extent_root;
10377
10378         btrfs_set_log_full_commit(root->fs_info, trans);
10379
10380         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10381         if (!cache)
10382                 return -ENOMEM;
10383
10384         btrfs_set_block_group_used(&cache->item, bytes_used);
10385         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10386         btrfs_set_block_group_flags(&cache->item, type);
10387
10388         cache->flags = type;
10389         cache->last_byte_to_unpin = (u64)-1;
10390         cache->cached = BTRFS_CACHE_FINISHED;
10391         cache->needs_free_space = 1;
10392         ret = exclude_super_stripes(root, cache);
10393         if (ret) {
10394                 /*
10395                  * We may have excluded something, so call this just in
10396                  * case.
10397                  */
10398                 free_excluded_extents(root, cache);
10399                 btrfs_put_block_group(cache);
10400                 return ret;
10401         }
10402
10403         add_new_free_space(cache, root->fs_info, chunk_offset,
10404                            chunk_offset + size);
10405
10406         free_excluded_extents(root, cache);
10407
10408 #ifdef CONFIG_BTRFS_DEBUG
10409         if (btrfs_should_fragment_free_space(root, cache)) {
10410                 u64 new_bytes_used = size - bytes_used;
10411
10412                 bytes_used += new_bytes_used >> 1;
10413                 fragment_free_space(root, cache);
10414         }
10415 #endif
10416         /*
10417          * Call to ensure the corresponding space_info object is created and
10418          * assigned to our block group, but don't update its counters just yet.
10419          * We want our bg to be added to the rbtree with its ->space_info set.
10420          */
10421         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10422                                 &cache->space_info);
10423         if (ret) {
10424                 btrfs_remove_free_space_cache(cache);
10425                 btrfs_put_block_group(cache);
10426                 return ret;
10427         }
10428
10429         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10430         if (ret) {
10431                 btrfs_remove_free_space_cache(cache);
10432                 btrfs_put_block_group(cache);
10433                 return ret;
10434         }
10435
10436         /*
10437          * Now that our block group has its ->space_info set and is inserted in
10438          * the rbtree, update the space info's counters.
10439          */
10440         trace_btrfs_add_block_group(root->fs_info, cache, 1);
10441         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10442                                 cache->bytes_super, &cache->space_info);
10443         if (ret) {
10444                 btrfs_remove_free_space_cache(cache);
10445                 spin_lock(&root->fs_info->block_group_cache_lock);
10446                 rb_erase(&cache->cache_node,
10447                          &root->fs_info->block_group_cache_tree);
10448                 RB_CLEAR_NODE(&cache->cache_node);
10449                 spin_unlock(&root->fs_info->block_group_cache_lock);
10450                 btrfs_put_block_group(cache);
10451                 return ret;
10452         }
10453         update_global_block_rsv(root->fs_info);
10454
10455         __link_block_group(cache->space_info, cache);
10456
10457         list_add_tail(&cache->bg_list, &trans->new_bgs);
10458
10459         set_avail_alloc_bits(extent_root->fs_info, type);
10460         return 0;
10461 }
10462
10463 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10464 {
10465         u64 extra_flags = chunk_to_extended(flags) &
10466                                 BTRFS_EXTENDED_PROFILE_MASK;
10467
10468         write_seqlock(&fs_info->profiles_lock);
10469         if (flags & BTRFS_BLOCK_GROUP_DATA)
10470                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10471         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10472                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10473         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10474                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10475         write_sequnlock(&fs_info->profiles_lock);
10476 }
10477
10478 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10479                              struct btrfs_root *root, u64 group_start,
10480                              struct extent_map *em)
10481 {
10482         struct btrfs_path *path;
10483         struct btrfs_block_group_cache *block_group;
10484         struct btrfs_free_cluster *cluster;
10485         struct btrfs_root *tree_root = root->fs_info->tree_root;
10486         struct btrfs_key key;
10487         struct inode *inode;
10488         struct kobject *kobj = NULL;
10489         int ret;
10490         int index;
10491         int factor;
10492         struct btrfs_caching_control *caching_ctl = NULL;
10493         bool remove_em;
10494
10495         root = root->fs_info->extent_root;
10496
10497         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10498         BUG_ON(!block_group);
10499         BUG_ON(!block_group->ro);
10500
10501         /*
10502          * Free the reserved super bytes from this block group before
10503          * remove it.
10504          */
10505         free_excluded_extents(root, block_group);
10506
10507         memcpy(&key, &block_group->key, sizeof(key));
10508         index = get_block_group_index(block_group);
10509         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10510                                   BTRFS_BLOCK_GROUP_RAID1 |
10511                                   BTRFS_BLOCK_GROUP_RAID10))
10512                 factor = 2;
10513         else
10514                 factor = 1;
10515
10516         /* make sure this block group isn't part of an allocation cluster */
10517         cluster = &root->fs_info->data_alloc_cluster;
10518         spin_lock(&cluster->refill_lock);
10519         btrfs_return_cluster_to_free_space(block_group, cluster);
10520         spin_unlock(&cluster->refill_lock);
10521
10522         /*
10523          * make sure this block group isn't part of a metadata
10524          * allocation cluster
10525          */
10526         cluster = &root->fs_info->meta_alloc_cluster;
10527         spin_lock(&cluster->refill_lock);
10528         btrfs_return_cluster_to_free_space(block_group, cluster);
10529         spin_unlock(&cluster->refill_lock);
10530
10531         path = btrfs_alloc_path();
10532         if (!path) {
10533                 ret = -ENOMEM;
10534                 goto out;
10535         }
10536
10537         /*
10538          * get the inode first so any iput calls done for the io_list
10539          * aren't the final iput (no unlinks allowed now)
10540          */
10541         inode = lookup_free_space_inode(tree_root, block_group, path);
10542
10543         mutex_lock(&trans->transaction->cache_write_mutex);
10544         /*
10545          * make sure our free spache cache IO is done before remove the
10546          * free space inode
10547          */
10548         spin_lock(&trans->transaction->dirty_bgs_lock);
10549         if (!list_empty(&block_group->io_list)) {
10550                 list_del_init(&block_group->io_list);
10551
10552                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10553
10554                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10555                 btrfs_wait_cache_io(root, trans, block_group,
10556                                     &block_group->io_ctl, path,
10557                                     block_group->key.objectid);
10558                 btrfs_put_block_group(block_group);
10559                 spin_lock(&trans->transaction->dirty_bgs_lock);
10560         }
10561
10562         if (!list_empty(&block_group->dirty_list)) {
10563                 list_del_init(&block_group->dirty_list);
10564                 btrfs_put_block_group(block_group);
10565         }
10566         spin_unlock(&trans->transaction->dirty_bgs_lock);
10567         mutex_unlock(&trans->transaction->cache_write_mutex);
10568
10569         if (!IS_ERR(inode)) {
10570                 ret = btrfs_orphan_add(trans, inode);
10571                 if (ret) {
10572                         btrfs_add_delayed_iput(inode);
10573                         goto out;
10574                 }
10575                 clear_nlink(inode);
10576                 /* One for the block groups ref */
10577                 spin_lock(&block_group->lock);
10578                 if (block_group->iref) {
10579                         block_group->iref = 0;
10580                         block_group->inode = NULL;
10581                         spin_unlock(&block_group->lock);
10582                         iput(inode);
10583                 } else {
10584                         spin_unlock(&block_group->lock);
10585                 }
10586                 /* One for our lookup ref */
10587                 btrfs_add_delayed_iput(inode);
10588         }
10589
10590         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10591         key.offset = block_group->key.objectid;
10592         key.type = 0;
10593
10594         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10595         if (ret < 0)
10596                 goto out;
10597         if (ret > 0)
10598                 btrfs_release_path(path);
10599         if (ret == 0) {
10600                 ret = btrfs_del_item(trans, tree_root, path);
10601                 if (ret)
10602                         goto out;
10603                 btrfs_release_path(path);
10604         }
10605
10606         spin_lock(&root->fs_info->block_group_cache_lock);
10607         rb_erase(&block_group->cache_node,
10608                  &root->fs_info->block_group_cache_tree);
10609         RB_CLEAR_NODE(&block_group->cache_node);
10610
10611         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10612                 root->fs_info->first_logical_byte = (u64)-1;
10613         spin_unlock(&root->fs_info->block_group_cache_lock);
10614
10615         down_write(&block_group->space_info->groups_sem);
10616         /*
10617          * we must use list_del_init so people can check to see if they
10618          * are still on the list after taking the semaphore
10619          */
10620         list_del_init(&block_group->list);
10621         if (list_empty(&block_group->space_info->block_groups[index])) {
10622                 kobj = block_group->space_info->block_group_kobjs[index];
10623                 block_group->space_info->block_group_kobjs[index] = NULL;
10624                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10625         }
10626         up_write(&block_group->space_info->groups_sem);
10627         if (kobj) {
10628                 kobject_del(kobj);
10629                 kobject_put(kobj);
10630         }
10631
10632         if (block_group->has_caching_ctl)
10633                 caching_ctl = get_caching_control(block_group);
10634         if (block_group->cached == BTRFS_CACHE_STARTED)
10635                 wait_block_group_cache_done(block_group);
10636         if (block_group->has_caching_ctl) {
10637                 down_write(&root->fs_info->commit_root_sem);
10638                 if (!caching_ctl) {
10639                         struct btrfs_caching_control *ctl;
10640
10641                         list_for_each_entry(ctl,
10642                                     &root->fs_info->caching_block_groups, list)
10643                                 if (ctl->block_group == block_group) {
10644                                         caching_ctl = ctl;
10645                                         atomic_inc(&caching_ctl->count);
10646                                         break;
10647                                 }
10648                 }
10649                 if (caching_ctl)
10650                         list_del_init(&caching_ctl->list);
10651                 up_write(&root->fs_info->commit_root_sem);
10652                 if (caching_ctl) {
10653                         /* Once for the caching bgs list and once for us. */
10654                         put_caching_control(caching_ctl);
10655                         put_caching_control(caching_ctl);
10656                 }
10657         }
10658
10659         spin_lock(&trans->transaction->dirty_bgs_lock);
10660         if (!list_empty(&block_group->dirty_list)) {
10661                 WARN_ON(1);
10662         }
10663         if (!list_empty(&block_group->io_list)) {
10664                 WARN_ON(1);
10665         }
10666         spin_unlock(&trans->transaction->dirty_bgs_lock);
10667         btrfs_remove_free_space_cache(block_group);
10668
10669         spin_lock(&block_group->space_info->lock);
10670         list_del_init(&block_group->ro_list);
10671
10672         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
10673                 WARN_ON(block_group->space_info->total_bytes
10674                         < block_group->key.offset);
10675                 WARN_ON(block_group->space_info->bytes_readonly
10676                         < block_group->key.offset);
10677                 WARN_ON(block_group->space_info->disk_total
10678                         < block_group->key.offset * factor);
10679         }
10680         block_group->space_info->total_bytes -= block_group->key.offset;
10681         block_group->space_info->bytes_readonly -= block_group->key.offset;
10682         block_group->space_info->disk_total -= block_group->key.offset * factor;
10683
10684         spin_unlock(&block_group->space_info->lock);
10685
10686         memcpy(&key, &block_group->key, sizeof(key));
10687
10688         lock_chunks(root);
10689         if (!list_empty(&em->list)) {
10690                 /* We're in the transaction->pending_chunks list. */
10691                 free_extent_map(em);
10692         }
10693         spin_lock(&block_group->lock);
10694         block_group->removed = 1;
10695         /*
10696          * At this point trimming can't start on this block group, because we
10697          * removed the block group from the tree fs_info->block_group_cache_tree
10698          * so no one can't find it anymore and even if someone already got this
10699          * block group before we removed it from the rbtree, they have already
10700          * incremented block_group->trimming - if they didn't, they won't find
10701          * any free space entries because we already removed them all when we
10702          * called btrfs_remove_free_space_cache().
10703          *
10704          * And we must not remove the extent map from the fs_info->mapping_tree
10705          * to prevent the same logical address range and physical device space
10706          * ranges from being reused for a new block group. This is because our
10707          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10708          * completely transactionless, so while it is trimming a range the
10709          * currently running transaction might finish and a new one start,
10710          * allowing for new block groups to be created that can reuse the same
10711          * physical device locations unless we take this special care.
10712          *
10713          * There may also be an implicit trim operation if the file system
10714          * is mounted with -odiscard. The same protections must remain
10715          * in place until the extents have been discarded completely when
10716          * the transaction commit has completed.
10717          */
10718         remove_em = (atomic_read(&block_group->trimming) == 0);
10719         /*
10720          * Make sure a trimmer task always sees the em in the pinned_chunks list
10721          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10722          * before checking block_group->removed).
10723          */
10724         if (!remove_em) {
10725                 /*
10726                  * Our em might be in trans->transaction->pending_chunks which
10727                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10728                  * and so is the fs_info->pinned_chunks list.
10729                  *
10730                  * So at this point we must be holding the chunk_mutex to avoid
10731                  * any races with chunk allocation (more specifically at
10732                  * volumes.c:contains_pending_extent()), to ensure it always
10733                  * sees the em, either in the pending_chunks list or in the
10734                  * pinned_chunks list.
10735                  */
10736                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10737         }
10738         spin_unlock(&block_group->lock);
10739
10740         if (remove_em) {
10741                 struct extent_map_tree *em_tree;
10742
10743                 em_tree = &root->fs_info->mapping_tree.map_tree;
10744                 write_lock(&em_tree->lock);
10745                 /*
10746                  * The em might be in the pending_chunks list, so make sure the
10747                  * chunk mutex is locked, since remove_extent_mapping() will
10748                  * delete us from that list.
10749                  */
10750                 remove_extent_mapping(em_tree, em);
10751                 write_unlock(&em_tree->lock);
10752                 /* once for the tree */
10753                 free_extent_map(em);
10754         }
10755
10756         unlock_chunks(root);
10757
10758         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10759         if (ret)
10760                 goto out;
10761
10762         btrfs_put_block_group(block_group);
10763         btrfs_put_block_group(block_group);
10764
10765         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10766         if (ret > 0)
10767                 ret = -EIO;
10768         if (ret < 0)
10769                 goto out;
10770
10771         ret = btrfs_del_item(trans, root, path);
10772 out:
10773         btrfs_free_path(path);
10774         return ret;
10775 }
10776
10777 struct btrfs_trans_handle *
10778 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10779                                      const u64 chunk_offset)
10780 {
10781         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10782         struct extent_map *em;
10783         struct map_lookup *map;
10784         unsigned int num_items;
10785
10786         read_lock(&em_tree->lock);
10787         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10788         read_unlock(&em_tree->lock);
10789         ASSERT(em && em->start == chunk_offset);
10790
10791         /*
10792          * We need to reserve 3 + N units from the metadata space info in order
10793          * to remove a block group (done at btrfs_remove_chunk() and at
10794          * btrfs_remove_block_group()), which are used for:
10795          *
10796          * 1 unit for adding the free space inode's orphan (located in the tree
10797          * of tree roots).
10798          * 1 unit for deleting the block group item (located in the extent
10799          * tree).
10800          * 1 unit for deleting the free space item (located in tree of tree
10801          * roots).
10802          * N units for deleting N device extent items corresponding to each
10803          * stripe (located in the device tree).
10804          *
10805          * In order to remove a block group we also need to reserve units in the
10806          * system space info in order to update the chunk tree (update one or
10807          * more device items and remove one chunk item), but this is done at
10808          * btrfs_remove_chunk() through a call to check_system_chunk().
10809          */
10810         map = em->map_lookup;
10811         num_items = 3 + map->num_stripes;
10812         free_extent_map(em);
10813
10814         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10815                                                            num_items, 1);
10816 }
10817
10818 /*
10819  * Process the unused_bgs list and remove any that don't have any allocated
10820  * space inside of them.
10821  */
10822 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10823 {
10824         struct btrfs_block_group_cache *block_group;
10825         struct btrfs_space_info *space_info;
10826         struct btrfs_root *root = fs_info->extent_root;
10827         struct btrfs_trans_handle *trans;
10828         int ret = 0;
10829
10830         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10831                 return;
10832
10833         spin_lock(&fs_info->unused_bgs_lock);
10834         while (!list_empty(&fs_info->unused_bgs)) {
10835                 u64 start, end;
10836                 int trimming;
10837
10838                 block_group = list_first_entry(&fs_info->unused_bgs,
10839                                                struct btrfs_block_group_cache,
10840                                                bg_list);
10841                 list_del_init(&block_group->bg_list);
10842
10843                 space_info = block_group->space_info;
10844
10845                 if (ret || btrfs_mixed_space_info(space_info)) {
10846                         btrfs_put_block_group(block_group);
10847                         continue;
10848                 }
10849                 spin_unlock(&fs_info->unused_bgs_lock);
10850
10851                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10852
10853                 /* Don't want to race with allocators so take the groups_sem */
10854                 down_write(&space_info->groups_sem);
10855                 spin_lock(&block_group->lock);
10856                 if (block_group->reserved ||
10857                     btrfs_block_group_used(&block_group->item) ||
10858                     block_group->ro ||
10859                     list_is_singular(&block_group->list)) {
10860                         /*
10861                          * We want to bail if we made new allocations or have
10862                          * outstanding allocations in this block group.  We do
10863                          * the ro check in case balance is currently acting on
10864                          * this block group.
10865                          */
10866                         spin_unlock(&block_group->lock);
10867                         up_write(&space_info->groups_sem);
10868                         goto next;
10869                 }
10870                 spin_unlock(&block_group->lock);
10871
10872                 /* We don't want to force the issue, only flip if it's ok. */
10873                 ret = inc_block_group_ro(block_group, 0);
10874                 up_write(&space_info->groups_sem);
10875                 if (ret < 0) {
10876                         ret = 0;
10877                         goto next;
10878                 }
10879
10880                 /*
10881                  * Want to do this before we do anything else so we can recover
10882                  * properly if we fail to join the transaction.
10883                  */
10884                 trans = btrfs_start_trans_remove_block_group(fs_info,
10885                                                      block_group->key.objectid);
10886                 if (IS_ERR(trans)) {
10887                         btrfs_dec_block_group_ro(root, block_group);
10888                         ret = PTR_ERR(trans);
10889                         goto next;
10890                 }
10891
10892                 /*
10893                  * We could have pending pinned extents for this block group,
10894                  * just delete them, we don't care about them anymore.
10895                  */
10896                 start = block_group->key.objectid;
10897                 end = start + block_group->key.offset - 1;
10898                 /*
10899                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10900                  * btrfs_finish_extent_commit(). If we are at transaction N,
10901                  * another task might be running finish_extent_commit() for the
10902                  * previous transaction N - 1, and have seen a range belonging
10903                  * to the block group in freed_extents[] before we were able to
10904                  * clear the whole block group range from freed_extents[]. This
10905                  * means that task can lookup for the block group after we
10906                  * unpinned it from freed_extents[] and removed it, leading to
10907                  * a BUG_ON() at btrfs_unpin_extent_range().
10908                  */
10909                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10910                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10911                                   EXTENT_DIRTY);
10912                 if (ret) {
10913                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10914                         btrfs_dec_block_group_ro(root, block_group);
10915                         goto end_trans;
10916                 }
10917                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10918                                   EXTENT_DIRTY);
10919                 if (ret) {
10920                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10921                         btrfs_dec_block_group_ro(root, block_group);
10922                         goto end_trans;
10923                 }
10924                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10925
10926                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10927                 spin_lock(&space_info->lock);
10928                 spin_lock(&block_group->lock);
10929
10930                 space_info->bytes_pinned -= block_group->pinned;
10931                 space_info->bytes_readonly += block_group->pinned;
10932                 percpu_counter_add(&space_info->total_bytes_pinned,
10933                                    -block_group->pinned);
10934                 block_group->pinned = 0;
10935
10936                 spin_unlock(&block_group->lock);
10937                 spin_unlock(&space_info->lock);
10938
10939                 /* DISCARD can flip during remount */
10940                 trimming = btrfs_test_opt(root->fs_info, DISCARD);
10941
10942                 /* Implicit trim during transaction commit. */
10943                 if (trimming)
10944                         btrfs_get_block_group_trimming(block_group);
10945
10946                 /*
10947                  * Btrfs_remove_chunk will abort the transaction if things go
10948                  * horribly wrong.
10949                  */
10950                 ret = btrfs_remove_chunk(trans, root,
10951                                          block_group->key.objectid);
10952
10953                 if (ret) {
10954                         if (trimming)
10955                                 btrfs_put_block_group_trimming(block_group);
10956                         goto end_trans;
10957                 }
10958
10959                 /*
10960                  * If we're not mounted with -odiscard, we can just forget
10961                  * about this block group. Otherwise we'll need to wait
10962                  * until transaction commit to do the actual discard.
10963                  */
10964                 if (trimming) {
10965                         spin_lock(&fs_info->unused_bgs_lock);
10966                         /*
10967                          * A concurrent scrub might have added us to the list
10968                          * fs_info->unused_bgs, so use a list_move operation
10969                          * to add the block group to the deleted_bgs list.
10970                          */
10971                         list_move(&block_group->bg_list,
10972                                   &trans->transaction->deleted_bgs);
10973                         spin_unlock(&fs_info->unused_bgs_lock);
10974                         btrfs_get_block_group(block_group);
10975                 }
10976 end_trans:
10977                 btrfs_end_transaction(trans, root);
10978 next:
10979                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10980                 btrfs_put_block_group(block_group);
10981                 spin_lock(&fs_info->unused_bgs_lock);
10982         }
10983         spin_unlock(&fs_info->unused_bgs_lock);
10984 }
10985
10986 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10987 {
10988         struct btrfs_space_info *space_info;
10989         struct btrfs_super_block *disk_super;
10990         u64 features;
10991         u64 flags;
10992         int mixed = 0;
10993         int ret;
10994
10995         disk_super = fs_info->super_copy;
10996         if (!btrfs_super_root(disk_super))
10997                 return -EINVAL;
10998
10999         features = btrfs_super_incompat_flags(disk_super);
11000         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11001                 mixed = 1;
11002
11003         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11004         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11005         if (ret)
11006                 goto out;
11007
11008         if (mixed) {
11009                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11010                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11011         } else {
11012                 flags = BTRFS_BLOCK_GROUP_METADATA;
11013                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11014                 if (ret)
11015                         goto out;
11016
11017                 flags = BTRFS_BLOCK_GROUP_DATA;
11018                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11019         }
11020 out:
11021         return ret;
11022 }
11023
11024 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
11025 {
11026         return unpin_extent_range(root, start, end, false);
11027 }
11028
11029 /*
11030  * It used to be that old block groups would be left around forever.
11031  * Iterating over them would be enough to trim unused space.  Since we
11032  * now automatically remove them, we also need to iterate over unallocated
11033  * space.
11034  *
11035  * We don't want a transaction for this since the discard may take a
11036  * substantial amount of time.  We don't require that a transaction be
11037  * running, but we do need to take a running transaction into account
11038  * to ensure that we're not discarding chunks that were released in
11039  * the current transaction.
11040  *
11041  * Holding the chunks lock will prevent other threads from allocating
11042  * or releasing chunks, but it won't prevent a running transaction
11043  * from committing and releasing the memory that the pending chunks
11044  * list head uses.  For that, we need to take a reference to the
11045  * transaction.
11046  */
11047 static int btrfs_trim_free_extents(struct btrfs_device *device,
11048                                    u64 minlen, u64 *trimmed)
11049 {
11050         u64 start = 0, len = 0;
11051         int ret;
11052
11053         *trimmed = 0;
11054
11055         /* Not writeable = nothing to do. */
11056         if (!device->writeable)
11057                 return 0;
11058
11059         /* No free space = nothing to do. */
11060         if (device->total_bytes <= device->bytes_used)
11061                 return 0;
11062
11063         ret = 0;
11064
11065         while (1) {
11066                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
11067                 struct btrfs_transaction *trans;
11068                 u64 bytes;
11069
11070                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11071                 if (ret)
11072                         return ret;
11073
11074                 down_read(&fs_info->commit_root_sem);
11075
11076                 spin_lock(&fs_info->trans_lock);
11077                 trans = fs_info->running_transaction;
11078                 if (trans)
11079                         atomic_inc(&trans->use_count);
11080                 spin_unlock(&fs_info->trans_lock);
11081
11082                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11083                                                  &start, &len);
11084                 if (trans)
11085                         btrfs_put_transaction(trans);
11086
11087                 if (ret) {
11088                         up_read(&fs_info->commit_root_sem);
11089                         mutex_unlock(&fs_info->chunk_mutex);
11090                         if (ret == -ENOSPC)
11091                                 ret = 0;
11092                         break;
11093                 }
11094
11095                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11096                 up_read(&fs_info->commit_root_sem);
11097                 mutex_unlock(&fs_info->chunk_mutex);
11098
11099                 if (ret)
11100                         break;
11101
11102                 start += len;
11103                 *trimmed += bytes;
11104
11105                 if (fatal_signal_pending(current)) {
11106                         ret = -ERESTARTSYS;
11107                         break;
11108                 }
11109
11110                 cond_resched();
11111         }
11112
11113         return ret;
11114 }
11115
11116 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
11117 {
11118         struct btrfs_fs_info *fs_info = root->fs_info;
11119         struct btrfs_block_group_cache *cache = NULL;
11120         struct btrfs_device *device;
11121         struct list_head *devices;
11122         u64 group_trimmed;
11123         u64 start;
11124         u64 end;
11125         u64 trimmed = 0;
11126         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11127         int ret = 0;
11128
11129         /*
11130          * try to trim all FS space, our block group may start from non-zero.
11131          */
11132         if (range->len == total_bytes)
11133                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11134         else
11135                 cache = btrfs_lookup_block_group(fs_info, range->start);
11136
11137         while (cache) {
11138                 if (cache->key.objectid >= (range->start + range->len)) {
11139                         btrfs_put_block_group(cache);
11140                         break;
11141                 }
11142
11143                 start = max(range->start, cache->key.objectid);
11144                 end = min(range->start + range->len,
11145                                 cache->key.objectid + cache->key.offset);
11146
11147                 if (end - start >= range->minlen) {
11148                         if (!block_group_cache_done(cache)) {
11149                                 ret = cache_block_group(cache, 0);
11150                                 if (ret) {
11151                                         btrfs_put_block_group(cache);
11152                                         break;
11153                                 }
11154                                 ret = wait_block_group_cache_done(cache);
11155                                 if (ret) {
11156                                         btrfs_put_block_group(cache);
11157                                         break;
11158                                 }
11159                         }
11160                         ret = btrfs_trim_block_group(cache,
11161                                                      &group_trimmed,
11162                                                      start,
11163                                                      end,
11164                                                      range->minlen);
11165
11166                         trimmed += group_trimmed;
11167                         if (ret) {
11168                                 btrfs_put_block_group(cache);
11169                                 break;
11170                         }
11171                 }
11172
11173                 cache = next_block_group(fs_info->tree_root, cache);
11174         }
11175
11176         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
11177         devices = &root->fs_info->fs_devices->alloc_list;
11178         list_for_each_entry(device, devices, dev_alloc_list) {
11179                 ret = btrfs_trim_free_extents(device, range->minlen,
11180                                               &group_trimmed);
11181                 if (ret)
11182                         break;
11183
11184                 trimmed += group_trimmed;
11185         }
11186         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
11187
11188         range->len = trimmed;
11189         return ret;
11190 }
11191
11192 /*
11193  * btrfs_{start,end}_write_no_snapshoting() are similar to
11194  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11195  * data into the page cache through nocow before the subvolume is snapshoted,
11196  * but flush the data into disk after the snapshot creation, or to prevent
11197  * operations while snapshoting is ongoing and that cause the snapshot to be
11198  * inconsistent (writes followed by expanding truncates for example).
11199  */
11200 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
11201 {
11202         percpu_counter_dec(&root->subv_writers->counter);
11203         /*
11204          * Make sure counter is updated before we wake up waiters.
11205          */
11206         smp_mb();
11207         if (waitqueue_active(&root->subv_writers->wait))
11208                 wake_up(&root->subv_writers->wait);
11209 }
11210
11211 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
11212 {
11213         if (atomic_read(&root->will_be_snapshoted))
11214                 return 0;
11215
11216         percpu_counter_inc(&root->subv_writers->counter);
11217         /*
11218          * Make sure counter is updated before we check for snapshot creation.
11219          */
11220         smp_mb();
11221         if (atomic_read(&root->will_be_snapshoted)) {
11222                 btrfs_end_write_no_snapshoting(root);
11223                 return 0;
11224         }
11225         return 1;
11226 }
11227
11228 static int wait_snapshoting_atomic_t(atomic_t *a)
11229 {
11230         schedule();
11231         return 0;
11232 }
11233
11234 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11235 {
11236         while (true) {
11237                 int ret;
11238
11239                 ret = btrfs_start_write_no_snapshoting(root);
11240                 if (ret)
11241                         break;
11242                 wait_on_atomic_t(&root->will_be_snapshoted,
11243                                  wait_snapshoting_atomic_t,
11244                                  TASK_UNINTERRUPTIBLE);
11245         }
11246 }